Compiling ipu-codegen v0.1.0 (/srv/home/gc-sdk/ipu-stack/crates/ipu-codegen) warning: unused import: `region_program` --> crates/ipu-codegen/src/estimate/mod.rs:22:23 | 22 | operation_cycles, region_program, | ^^^^^^^^^^^^^^ | = note: `#[warn(unused_imports)]` (part of `#[warn(unused)]`) on by default warning: unused import: `memory::*` --> crates/ipu-codegen/src/estimate/mod.rs:39:16 | 39 | pub(crate) use memory::*; | ^^^^^^^^^ warning: unused import: `contiguous_axis_blocks` --> crates/ipu-codegen/src/storage/mod.rs:12:49 | 12 | ByteTraversal, StridedSpan, byte_traversal, contiguous_axis_blocks, contiguous_axis_stride, panel_byte_traversal, | ^^^^^^^^^^^^^^^^^^^^^^ warning: unused imports: `axis_tiling` and `broadcast_operand_tiling` --> crates/ipu-codegen/src/tensor/mod.rs:99:5 | 99 | axis_tiling, broadcast_operand_tiling, project_tiling, same_distribution, | ^^^^^^^^^^^ ^^^^^^^^^^^^^^^^^^^^^^^^ warning: methods `objectives` and `total_with_exchange` are never used --> crates/ipu-codegen/src/estimate/memory.rs:57:19 | 46 | impl MemoryPeaks { | ---------------- methods in this implementation ... 57 | pub(crate) fn objectives(self) -> [u64; 5] { | ^^^^^^^^^^ ... 67 | pub(crate) fn total_with_exchange(self) -> u64 { | ^^^^^^^^^^^^^^^^^^^ | = note: `#[warn(dead_code)]` (part of `#[warn(unused)]`) on by default warning: function `tensor_memory` is never used --> crates/ipu-codegen/src/estimate/memory.rs:106:15 | 106 | pub(crate) fn tensor_memory(tensor: &TensorType) -> MemoryUsage { | ^^^^^^^^^^^^^ warning: function `operation_cycles` is never used --> crates/ipu-codegen/src/estimate/mid.rs:26:15 | 26 | pub(crate) fn operation_cycles<'a>( | ^^^^^^^^^^^^^^^^ warning: function `maximum_axis_shard_extent` is never used --> crates/ipu-codegen/src/estimate/tensor.rs:25:15 | 25 | pub(crate) fn maximum_axis_shard_extent(tensor: &TensorType, axis: usize) -> u64 { | ^^^^^^^^^^^^^^^^^^^^^^^^^ warning: constant `WORKER_CALL_CYCLES` is never used --> crates/ipu-codegen/src/kernel/copy.rs:18:18 | 18 | pub(crate) const WORKER_CALL_CYCLES: u64 = ipu_target::ipu21::costs::COSTS.local_copy_call_cycles; | ^^^^^^^^^^^^^^^^^^ warning: field `complete_rows` is never read --> crates/ipu-codegen/src/kernel/binding.rs:326:9 | 323 | pub(crate) struct OutputCapability { | ---------------- field in this struct ... 326 | pub complete_rows: bool, | ^^^^^^^^^^^^^ warning: function `supports_row_major_population` is never used --> crates/ipu-codegen/src/kernel/rearrange.rs:307:15 | 307 | pub(crate) fn supports_row_major_population(order: ElementOrder) -> bool { | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ warning: function `is_single_use` is never used --> crates/ipu-codegen/src/mid/rewrite.rs:44:15 | 44 | pub(crate) fn is_single_use( | ^^^^^^^^^^^^^ warning: function `producer_through_copies` is never used --> crates/ipu-codegen/src/mid/rewrite.rs:141:15 | 141 | pub(crate) fn producer_through_copies( | ^^^^^^^^^^^^^^^^^^^^^^^ warning: function `fusion_pays` is never used --> crates/ipu-codegen/src/mid/rewrite.rs:171:15 | 171 | pub(crate) fn fusion_pays<'a>( | ^^^^^^^^^^^ warning: method `input_axis` is never used --> crates/ipu-codegen/src/tensor/mod.rs:54:19 | 38 | impl<'a> Broadcast<'a> { | ---------------------- method in this implementation ... 54 | pub(crate) fn input_axis(self, output_axis: usize) -> Option { | ^^^^^^^^^^ warning: method `minimum_extent` is never used --> crates/ipu-codegen/src/tensor/resolved.rs:50:19 | 15 | impl ResolvedAxis { | ----------------- method in this implementation ... 50 | pub(crate) fn minimum_extent(&self) -> u32 { | ^^^^^^^^^^^^^^ warning: function `broadcast_operand_tiling` is never used --> crates/ipu-codegen/src/tensor/resolved.rs:503:15 | 503 | pub(crate) fn broadcast_operand_tiling( | ^^^^^^^^^^^^^^^^^^^^^^^^ warning: function `axis_tiling` is never used --> crates/ipu-codegen/src/tensor/resolved.rs:514:15 | 514 | pub(crate) fn axis_tiling(tensor: &TensorType, axis: usize) -> Option<&AxisTiling> { | ^^^^^^^^^^^ warning: `ipu-codegen` (lib test) generated 18 warnings (run `cargo fix --lib -p ipu-codegen --tests` to apply 4 suggestions) Finished `release` profile [optimized] target(s) in 33.76s Running unittests src/lib.rs (target/release/deps/ipu_codegen-8a4acda8960e2f33) running 1 test COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=323127 MemoryPeaks { standard: 138264, interleaved: 232448, total: 305152, exchange_rows: 0, maximum_standard_allocation: 93312 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=323127 MemoryPeaks { standard: 138264, interleaved: 232448, total: 305152, exchange_rows: 0, maximum_standard_allocation: 93312 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=344071 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=718407 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741250 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=324212 MemoryPeaks { standard: 138264, interleaved: 232448, total: 305152, exchange_rows: 0, maximum_standard_allocation: 93312 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=324212 MemoryPeaks { standard: 138264, interleaved: 232448, total: 305152, exchange_rows: 0, maximum_standard_allocation: 93312 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297887 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=291874 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304313 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735543 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758386 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=345156 MemoryPeaks { standard: 174080, interleaved: 131072, total: 305152, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=293080 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315023 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=309010 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321449 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=275944 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=248016 MemoryPeaks { standard: 124288, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 93696 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=275944 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=248016 MemoryPeaks { standard: 117120, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 93696 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=293080 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=268264 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 96768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=169775 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 955840, interleaved: 214016, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=719487 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 955840, interleaved: 214016, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=741787 MemoryPeaks { standard: 948672, interleaved: 296960, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 218560, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=298429 MemoryPeaks { standard: 211392, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 165280, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=292959 MemoryPeaks { standard: 158112, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 229696, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=304855 MemoryPeaks { standard: 222528, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 955840, interleaved: 131072, total: 981504, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=736623 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1057280, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 955840, interleaved: 131072, total: 955840, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=758923 MemoryPeaks { standard: 1031616, interleaved: 131072, total: 1031616, exchange_rows: 0, maximum_standard_allocation: 737856 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294165 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 218560, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=315565 MemoryPeaks { standard: 294336, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 165280, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=310095 MemoryPeaks { standard: 241056, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 122976 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 229696, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=321991 MemoryPeaks { standard: 305472, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 105408 } winner=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=277029 MemoryPeaks { standard: 126552, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=249102 MemoryPeaks { standard: 124288, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 93696 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=277029 MemoryPeaks { standard: 119384, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=249102 MemoryPeaks { standard: 117120, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 93696 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=294165 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 92736 } winner=269350 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 96768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=285532 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289007 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142095 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139934 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143361 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=302668 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306143 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154856 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159231 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157070 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160497 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2664185 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=2632193 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=2765652 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=2733660 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=1000082 MemoryPeaks { standard: 851008, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=968090 MemoryPeaks { standard: 843840, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=973261 MemoryPeaks { standard: 637024, interleaved: 508928, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=941269 MemoryPeaks { standard: 629856, interleaved: 508928, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=1024466 MemoryPeaks { standard: 893632, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=992474 MemoryPeaks { standard: 886464, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917760 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=2681321 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2649329 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2782788 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2750796 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=1017218 MemoryPeaks { standard: 851008, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=985226 MemoryPeaks { standard: 926784, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=990397 MemoryPeaks { standard: 637024, interleaved: 425984, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=958405 MemoryPeaks { standard: 712800, interleaved: 425984, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1041602 MemoryPeaks { standard: 893632, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1009610 MemoryPeaks { standard: 969408, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2664185 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=2632193 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=2765652 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=2733660 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=899515 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=899515 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=867523 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=867523 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1000082 MemoryPeaks { standard: 851008, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=968090 MemoryPeaks { standard: 843840, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=973261 MemoryPeaks { standard: 637024, interleaved: 508928, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=941269 MemoryPeaks { standard: 629856, interleaved: 508928, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1024466 MemoryPeaks { standard: 893632, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=992474 MemoryPeaks { standard: 886464, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=752769 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=2681321 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2649329 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2782788 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2750796 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=351855 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=884659 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=884659 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1017218 MemoryPeaks { standard: 851008, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=985226 MemoryPeaks { standard: 926784, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=990397 MemoryPeaks { standard: 637024, interleaved: 425984, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=958405 MemoryPeaks { standard: 712800, interleaved: 425984, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1041602 MemoryPeaks { standard: 893632, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1009610 MemoryPeaks { standard: 969408, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=899515 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=821094 MemoryPeaks { standard: 473728, interleaved: 499712, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=867523 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=821094 MemoryPeaks { standard: 473728, interleaved: 499712, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=916651 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=836326 MemoryPeaks { standard: 473728, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=884659 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=836326 MemoryPeaks { standard: 473728, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=550835 MemoryPeaks { standard: 272128, interleaved: 262144, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=531800 MemoryPeaks { standard: 272128, interleaved: 354304, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 320 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=550835 MemoryPeaks { standard: 272128, interleaved: 262144, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=531800 MemoryPeaks { standard: 272128, interleaved: 354304, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 320 }) COARSE_DROP old=328316 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=328316 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=333519 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=333519 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158801 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=158801 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=154912 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=134192 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=159191 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=134192 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=345452 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=503461 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=345452 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=585651 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=585651 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=350655 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=585651 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=350655 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=170734 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=175937 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=585651 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=175937 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172048 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172048 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=176327 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=176327 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=148790 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_REPLACED old=153598 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=138857 MemoryPeaks { standard: 59768, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=153598 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=138857 MemoryPeaks { standard: 52600, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=933403 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=964343 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376702 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=367848 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=367848 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384004 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=384004 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=170690 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=950539 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=981479 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=364698 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393838 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=384984 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=384984 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=401140 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=401140 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=175132 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_REPLACED old=347562 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=314822 MemoryPeaks { standard: 162688, interleaved: 237568, total: 310976, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=347562 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=314822 MemoryPeaks { standard: 155520, interleaved: 311296, total: 377536, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=364698 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=330054 MemoryPeaks { standard: 162688, interleaved: 163840, total: 310976, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1367900 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416051 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=531952 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=531952 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519153 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=412435 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544414 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=399745 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1385036 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433187 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=501837 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549088 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=536289 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=561550 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=484701 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=443358 MemoryPeaks { standard: 236160, interleaved: 376832, total: 515648, exchange_rows: 0, maximum_standard_allocation: 188928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=484701 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=438126 MemoryPeaks { standard: 240768, interleaved: 303104, total: 446784, exchange_rows: 0, maximum_standard_allocation: 186880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=501837 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=453358 MemoryPeaks { standard: 240768, interleaved: 229376, total: 446784, exchange_rows: 0, maximum_standard_allocation: 186880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=238885 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=238885 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=238885 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=238885 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2665248 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2633256 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2766172 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2734180 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=1000622 MemoryPeaks { standard: 851008, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=968630 MemoryPeaks { standard: 843840, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=974344 MemoryPeaks { standard: 637024, interleaved: 508928, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=942352 MemoryPeaks { standard: 629856, interleaved: 508928, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=1025006 MemoryPeaks { standard: 893632, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=993014 MemoryPeaks { standard: 886464, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=918843 MemoryPeaks { standard: 525568, interleaved: 491520, total: 965248, exchange_rows: 0, maximum_standard_allocation: 414720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=2682384 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=2650392 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=2783308 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=2751316 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=1017758 MemoryPeaks { standard: 851008, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=985766 MemoryPeaks { standard: 926784, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=991480 MemoryPeaks { standard: 637024, interleaved: 425984, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=959488 MemoryPeaks { standard: 712800, interleaved: 425984, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1042142 MemoryPeaks { standard: 893632, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=1010150 MemoryPeaks { standard: 969408, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2665248 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2633256 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=2766172 MemoryPeaks { standard: 3786304, interleaved: 508928, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2734180 MemoryPeaks { standard: 3779136, interleaved: 508928, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=900598 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=900598 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=868606 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=868606 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=1000622 MemoryPeaks { standard: 851008, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=968630 MemoryPeaks { standard: 843840, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=974344 MemoryPeaks { standard: 637024, interleaved: 508928, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=440439 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=440439 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=942352 MemoryPeaks { standard: 629856, interleaved: 508928, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452050 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=189115 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=135662 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=199826 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=199826 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=194533 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194533 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=146237 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=201434 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=135662 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=457575 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=457575 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469186 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=355038 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363193 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=469186 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=206251 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=216962 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211669 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=218570 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169062 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_REPLACED old=189115 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=170631 MemoryPeaks { standard: 76288, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=164597 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=164597 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=169794 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=137658 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=372174 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380329 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_REPLACED old=206251 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=185863 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1025006 MemoryPeaks { standard: 893632, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=380329 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=178943 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=993014 MemoryPeaks { standard: 886464, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=753853 MemoryPeaks { standard: 514688, interleaved: 491520, total: 967168, exchange_rows: 0, maximum_standard_allocation: 429440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=2682384 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3792384, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=178943 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=181733 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=186930 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=147714 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=2650392 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3868160, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2783308 MemoryPeaks { standard: 3786304, interleaved: 425984, total: 3786304, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=2751316 MemoryPeaks { standard: 3862080, interleaved: 425984, total: 3862080, exchange_rows: 0, maximum_standard_allocation: 2939328 } winner=353536 MemoryPeaks { standard: 989184, interleaved: 10240, total: 989184, exchange_rows: 0, maximum_standard_allocation: 499840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 320 }) COARSE_DROP old=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=161807 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=144618 MemoryPeaks { standard: 61816, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=161807 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=144618 MemoryPeaks { standard: 54648, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=885742 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=885742 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1017758 MemoryPeaks { standard: 851008, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=985766 MemoryPeaks { standard: 926784, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=991480 MemoryPeaks { standard: 637024, interleaved: 425984, total: 993024, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=959488 MemoryPeaks { standard: 712800, interleaved: 425984, total: 1068800, exchange_rows: 0, maximum_standard_allocation: 489888 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1042142 MemoryPeaks { standard: 893632, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1010150 MemoryPeaks { standard: 969408, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=900598 MemoryPeaks { standard: 484096, interleaved: 508928, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=822177 MemoryPeaks { standard: 473728, interleaved: 499712, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=868606 MemoryPeaks { standard: 476928, interleaved: 508928, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=822177 MemoryPeaks { standard: 473728, interleaved: 499712, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=917734 MemoryPeaks { standard: 484096, interleaved: 425984, total: 857088, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=837409 MemoryPeaks { standard: 473728, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=885742 MemoryPeaks { standard: 559872, interleaved: 425984, total: 932864, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=837409 MemoryPeaks { standard: 473728, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 373248 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=239988 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=239988 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=239988 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=239988 MemoryPeaks { standard: 157696, interleaved: 65536, total: 223232, exchange_rows: 0, maximum_standard_allocation: 101376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=441522 MemoryPeaks { standard: 551488, interleaved: 148480, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=441522 MemoryPeaks { standard: 544320, interleaved: 231424, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 551488, interleaved: 148480, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=452590 MemoryPeaks { standard: 544320, interleaved: 231424, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=190484 MemoryPeaks { standard: 72576, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=137832 MemoryPeaks { standard: 71680, interleaved: 200704, total: 233472, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 132160, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=200652 MemoryPeaks { standard: 124992, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=195902 MemoryPeaks { standard: 97888, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195902 MemoryPeaks { standard: 90720, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 70560 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 134848, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=148407 MemoryPeaks { standard: 91136, interleaved: 98304, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=202260 MemoryPeaks { standard: 127680, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 60480 } winner=137832 MemoryPeaks { standard: 78848, interleaved: 149504, total: 189440, exchange_rows: 0, maximum_standard_allocation: 62720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=458658 MemoryPeaks { standard: 551488, interleaved: 65536, total: 556544, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=458658 MemoryPeaks { standard: 627264, interleaved: 65536, total: 632320, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 551488, interleaved: 65536, total: 551488, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=469726 MemoryPeaks { standard: 627264, interleaved: 65536, total: 627264, exchange_rows: 0, maximum_standard_allocation: 423360 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207620 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 132160, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=217788 MemoryPeaks { standard: 207936, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213038 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 134848, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=219396 MemoryPeaks { standard: 210624, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=190484 MemoryPeaks { standard: 79744, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56448 } winner=172208 MemoryPeaks { standard: 76288, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_REPLACED old=207620 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=187440 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=587537 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=587537 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=604152 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=604152 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252089 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=245860 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=255497 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=171249 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604673 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621288 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=253510 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269225 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=262996 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=272633 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=181824 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_REPLACED old=236374 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=215711 MemoryPeaks { standard: 105088, interleaved: 172032, total: 228352, exchange_rows: 0, maximum_standard_allocation: 78336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_REPLACED old=236374 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=215711 MemoryPeaks { standard: 97920, interleaved: 245760, total: 294912, exchange_rows: 0, maximum_standard_allocation: 78336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=329399 MemoryPeaks { standard: 385600, interleaved: 148480, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=306845 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=503052 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515668 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221847 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221847 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=329399 MemoryPeaks { standard: 378432, interleaved: 231424, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=311184 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=334059 MemoryPeaks { standard: 385600, interleaved: 148480, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=311184 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=334059 MemoryPeaks { standard: 378432, interleaved: 231424, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160188 MemoryPeaks { standard: 95296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=160188 MemoryPeaks { standard: 88128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=156842 MemoryPeaks { standard: 70240, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=216989 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=216989 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=224379 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=173024 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=520188 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532804 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=227267 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=150438 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=156842 MemoryPeaks { standard: 63072, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 49056 } winner=136340 MemoryPeaks { standard: 60736, interleaved: 206848, total: 223232, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160578 MemoryPeaks { standard: 95936, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=160578 MemoryPeaks { standard: 88768, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 42048 } winner=136340 MemoryPeaks { standard: 67904, interleaved: 136192, total: 159744, exchange_rows: 0, maximum_standard_allocation: 51392 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 385600, interleaved: 65536, total: 409088, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=346535 MemoryPeaks { standard: 461376, interleaved: 65536, total: 484864, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 385600, interleaved: 65536, total: 385600, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=351195 MemoryPeaks { standard: 461376, interleaved: 65536, total: 461376, exchange_rows: 0, maximum_standard_allocation: 294336 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=150438 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=227267 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=147413 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=145663 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=238983 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=234125 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=151266 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=142727 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=234125 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=241515 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=323981 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=323981 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328320 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328320 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=241515 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163767 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=161435 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=161435 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=167574 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_REPLACED old=210131 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=190980 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=167574 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=172664 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=172664 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_REPLACED old=210131 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=190980 MemoryPeaks { standard: 80640, interleaved: 245760, total: 278528, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=177324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=164549 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=164549 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=168402 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=173978 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=168402 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158591 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=177714 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=177714 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=150938 MemoryPeaks { standard: 94208, interleaved: 65536, total: 159744, exchange_rows: 0, maximum_standard_allocation: 70656 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 368 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_REPLACED old=155528 MemoryPeaks { standard: 62104, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=140932 MemoryPeaks { standard: 59768, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=155528 MemoryPeaks { standard: 54936, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40320 } winner=140932 MemoryPeaks { standard: 52600, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 639616, interleaved: 181248, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=501970 MemoryPeaks { standard: 632448, interleaved: 264192, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 639616, interleaved: 181248, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=515129 MemoryPeaks { standard: 632448, interleaved: 264192, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221305 MemoryPeaks { standard: 151168, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=221305 MemoryPeaks { standard: 144000, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 112576, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=215904 MemoryPeaks { standard: 105408, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 81984 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 155520, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=223837 MemoryPeaks { standard: 148352, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=171343 MemoryPeaks { standard: 98688, interleaved: 159744, total: 199680, exchange_rows: 0, maximum_standard_allocation: 77440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 639616, interleaved: 98304, total: 667648, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=519106 MemoryPeaks { standard: 715392, interleaved: 98304, total: 743424, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 639616, interleaved: 98304, total: 639616, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=532265 MemoryPeaks { standard: 715392, interleaved: 98304, total: 715392, exchange_rows: 0, maximum_standard_allocation: 491904 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=226182 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 151168, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=238441 MemoryPeaks { standard: 226944, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=233040 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 155520, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=240973 MemoryPeaks { standard: 231296, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=161775 MemoryPeaks { standard: 99328, interleaved: 98304, total: 191872, exchange_rows: 0, maximum_standard_allocation: 80640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_REPLACED old=209046 MemoryPeaks { standard: 92600, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=189680 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=209046 MemoryPeaks { standard: 85432, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 64512 } winner=189680 MemoryPeaks { standard: 80640, interleaved: 245760, total: 278528, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 764032, interleaved: 181248, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=588618 MemoryPeaks { standard: 756864, interleaved: 264192, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 764032, interleaved: 181248, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=604690 MemoryPeaks { standard: 756864, interleaved: 264192, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=252631 MemoryPeaks { standard: 178816, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=252631 MemoryPeaks { standard: 171648, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 133312, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=246945 MemoryPeaks { standard: 126144, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 184704, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=256039 MemoryPeaks { standard: 177536, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=172992 MemoryPeaks { standard: 109568, interleaved: 149504, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 764032, interleaved: 98304, total: 778240, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=605754 MemoryPeaks { standard: 839808, interleaved: 98304, total: 854016, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 764032, interleaved: 98304, total: 764032, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=621826 MemoryPeaks { standard: 839808, interleaved: 98304, total: 839808, exchange_rows: 0, maximum_standard_allocation: 588672 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=254595 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 178816, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=269767 MemoryPeaks { standard: 254592, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=264081 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 98112 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 184704, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_DROP old=273175 MemoryPeaks { standard: 260480, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=183567 MemoryPeaks { standard: 109568, interleaved: 98304, total: 201472, exchange_rows: 0, maximum_standard_allocation: 89600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 320 }) COARSE_REPLACED old=237459 MemoryPeaks { standard: 105664, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=216797 MemoryPeaks { standard: 105088, interleaved: 172032, total: 228352, exchange_rows: 0, maximum_standard_allocation: 78336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=237459 MemoryPeaks { standard: 98496, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 76608 } winner=216797 MemoryPeaks { standard: 97920, interleaved: 245760, total: 294912, exchange_rows: 0, maximum_standard_allocation: 78336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=356121 MemoryPeaks { standard: 427072, interleaved: 148480, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=356121 MemoryPeaks { standard: 419904, interleaved: 231424, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 427072, interleaved: 148480, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=363733 MemoryPeaks { standard: 419904, interleaved: 231424, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170304 MemoryPeaks { standard: 104512, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=170304 MemoryPeaks { standard: 97344, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 77152, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=166382 MemoryPeaks { standard: 69984, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 54432 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=171036 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=171036 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 46656 } winner=139778 MemoryPeaks { standard: 68320, interleaved: 130048, total: 153600, exchange_rows: 0, maximum_standard_allocation: 51744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 427072, interleaved: 65536, total: 445952, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=373257 MemoryPeaks { standard: 502848, interleaved: 65536, total: 521728, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 427072, interleaved: 65536, total: 427072, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=380869 MemoryPeaks { standard: 502848, interleaved: 65536, total: 502848, exchange_rows: 0, maximum_standard_allocation: 326592 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180728 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180728 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=187440 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=187440 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=183518 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=183518 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_DROP old=188172 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=149834 MemoryPeaks { standard: 88064, interleaved: 65536, total: 153600, exchange_rows: 0, maximum_standard_allocation: 64512 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 336 }) COARSE_REPLACED old=163592 MemoryPeaks { standard: 64408, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=146564 MemoryPeaks { standard: 61816, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=163592 MemoryPeaks { standard: 57240, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 44352 } winner=146564 MemoryPeaks { standard: 54648, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 41472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=396094 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=396094 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=404390 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183228 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=183228 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=180062 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=180062 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=143144 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=185322 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185322 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=161019 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=413230 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=413230 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=421526 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=191168 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=200364 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=197198 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=202458 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=155834 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_REPLACED old=174032 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=154908 MemoryPeaks { standard: 66048, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 47104 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=174032 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=154908 MemoryPeaks { standard: 58880, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 47104 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=395011 MemoryPeaks { standard: 484096, interleaved: 148480, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=395011 MemoryPeaks { standard: 476928, interleaved: 231424, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 484096, interleaved: 148480, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=403850 MemoryPeaks { standard: 476928, interleaved: 231424, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=182168 MemoryPeaks { standard: 115456, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=182168 MemoryPeaks { standard: 108288, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=178459 MemoryPeaks { standard: 86656, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=178459 MemoryPeaks { standard: 79488, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 61824 } winner=141074 MemoryPeaks { standard: 73728, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 56320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=184262 MemoryPeaks { standard: 119040, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184262 MemoryPeaks { standard: 111872, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 52992 } winner=158871 MemoryPeaks { standard: 89600, interleaved: 98304, total: 182752, exchange_rows: 0, maximum_standard_allocation: 72128 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=412147 MemoryPeaks { standard: 484096, interleaved: 65536, total: 496640, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=412147 MemoryPeaks { standard: 559872, interleaved: 65536, total: 572416, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 484096, interleaved: 65536, total: 484096, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=420986 MemoryPeaks { standard: 559872, interleaved: 65536, total: 559872, exchange_rows: 0, maximum_standard_allocation: 370944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=189565 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=199304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=195595 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=201398 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=153764 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_REPLACED old=172429 MemoryPeaks { standard: 69376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=153123 MemoryPeaks { standard: 66048, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 47104 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=172429 MemoryPeaks { standard: 62208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 48384 } winner=153123 MemoryPeaks { standard: 58880, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 47104 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=932325 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1266880, interleaved: 246784, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=963808 MemoryPeaks { standard: 1259712, interleaved: 329728, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=376160 MemoryPeaks { standard: 289984, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=376160 MemoryPeaks { standard: 282816, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 217120, interleaved: 246784, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=366763 MemoryPeaks { standard: 209952, interleaved: 329728, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383462 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=168570 MemoryPeaks { standard: 181216, interleaved: 185344, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1290752, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=949461 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1366528, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1266880, interleaved: 163840, total: 1266880, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=980944 MemoryPeaks { standard: 1342656, interleaved: 163840, total: 1342656, exchange_rows: 0, maximum_standard_allocation: 979776 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=363613 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 289984, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=393296 MemoryPeaks { standard: 365760, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 217120, interleaved: 163840, total: 357632, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=383899 MemoryPeaks { standard: 292896, interleaved: 163840, total: 433408, exchange_rows: 0, maximum_standard_allocation: 163296 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_DROP old=400598 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 139968 } winner=173012 MemoryPeaks { standard: 181216, interleaved: 163840, total: 340352, exchange_rows: 0, maximum_standard_allocation: 164640 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 32, output_columns: 336 }) COARSE_REPLACED old=346477 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=313737 MemoryPeaks { standard: 162688, interleaved: 237568, total: 310976, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=346477 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=313737 MemoryPeaks { standard: 155520, interleaved: 311296, total: 377536, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=363613 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 124992 } winner=328969 MemoryPeaks { standard: 162688, interleaved: 163840, total: 310976, exchange_rows: 0, maximum_standard_allocation: 124416 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=551919 MemoryPeaks { standard: 272128, interleaved: 262144, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=532884 MemoryPeaks { standard: 272128, interleaved: 354304, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=551919 MemoryPeaks { standard: 272128, interleaved: 262144, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=532884 MemoryPeaks { standard: 272128, interleaved: 354304, total: 504832, exchange_rows: 0, maximum_standard_allocation: 206080 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=504546 MemoryPeaks { standard: 269568, interleaved: 262144, total: 505472, exchange_rows: 0, maximum_standard_allocation: 209920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=586735 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=586735 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=586735 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=586735 MemoryPeaks { standard: 267864, interleaved: 229376, total: 468992, exchange_rows: 0, maximum_standard_allocation: 183168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1368974 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1899328, interleaved: 312320, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1416582 MemoryPeaks { standard: 1892160, interleaved: 395264, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=532493 MemoryPeaks { standard: 429376, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=532493 MemoryPeaks { standard: 422208, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 322528, interleaved: 312320, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=520237 MemoryPeaks { standard: 315360, interleaved: 395264, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 451008, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=413520 MemoryPeaks { standard: 265088, interleaved: 262144, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 320 }) COARSE_DROP old=544955 MemoryPeaks { standard: 443840, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=400830 MemoryPeaks { standard: 265088, interleaved: 323584, total: 507392, exchange_rows: 0, maximum_standard_allocation: 218240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 320 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1918464, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1386110 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1994240, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1899328, interleaved: 229376, total: 1899328, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1433718 MemoryPeaks { standard: 1975104, interleaved: 229376, total: 1975104, exchange_rows: 0, maximum_standard_allocation: 1471680 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=502921 MemoryPeaks { standard: 321408, interleaved: 229376, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 429376, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=549629 MemoryPeaks { standard: 505152, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 322528, interleaved: 229376, total: 516864, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=537373 MemoryPeaks { standard: 398304, interleaved: 229376, total: 592640, exchange_rows: 0, maximum_standard_allocation: 245280 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 451008, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=562091 MemoryPeaks { standard: 526784, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 210240 } winner=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=485785 MemoryPeaks { standard: 238464, interleaved: 395264, total: 524288, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=444443 MemoryPeaks { standard: 236160, interleaved: 376832, total: 515648, exchange_rows: 0, maximum_standard_allocation: 188928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=485785 MemoryPeaks { standard: 245632, interleaved: 312320, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=439211 MemoryPeaks { standard: 240768, interleaved: 303104, total: 446784, exchange_rows: 0, maximum_standard_allocation: 186880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=502921 MemoryPeaks { standard: 245632, interleaved: 229376, total: 448512, exchange_rows: 0, maximum_standard_allocation: 185472 } winner=454443 MemoryPeaks { standard: 240768, interleaved: 229376, total: 446784, exchange_rows: 0, maximum_standard_allocation: 186880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=171845 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 320 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 323392, interleaved: 148480, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=286616 MemoryPeaks { standard: 316224, interleaved: 231424, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 323392, interleaved: 148480, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289548 MemoryPeaks { standard: 316224, interleaved: 231424, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 79168, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=143700 MemoryPeaks { standard: 72000, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 59872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142082 MemoryPeaks { standard: 52704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 40992 } winner=139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 81344, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144966 MemoryPeaks { standard: 74176, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 35136 } winner=139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 323392, interleaved: 65536, total: 353792, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=303752 MemoryPeaks { standard: 399168, interleaved: 65536, total: 429568, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 323392, interleaved: 65536, total: 323392, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306684 MemoryPeaks { standard: 399168, interleaved: 65536, total: 399168, exchange_rows: 0, maximum_standard_allocation: 245952 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157004 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160836 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159218 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=162102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 354496, interleaved: 148480, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=307928 MemoryPeaks { standard: 347328, interleaved: 231424, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 354496, interleaved: 148480, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=311724 MemoryPeaks { standard: 347328, interleaved: 231424, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151944 MemoryPeaks { standard: 87232, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=151944 MemoryPeaks { standard: 80064, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=149462 MemoryPeaks { standard: 65056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 58368, interleaved: 147456, total: 187392, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=149462 MemoryPeaks { standard: 57888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 45024 } winner=147768 MemoryPeaks { standard: 51200, interleaved: 229376, total: 262144, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 320 }) COARSE_DROP old=152772 MemoryPeaks { standard: 88640, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=152772 MemoryPeaks { standard: 81472, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 38592 } winner=144884 MemoryPeaks { standard: 69568, interleaved: 142336, total: 165888, exchange_rows: 0, maximum_standard_allocation: 52800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 354496, interleaved: 65536, total: 381440, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=325064 MemoryPeaks { standard: 430272, interleaved: 65536, total: 457216, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 354496, interleaved: 65536, total: 354496, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=328860 MemoryPeaks { standard: 430272, interleaved: 65536, total: 430272, exchange_rows: 0, maximum_standard_allocation: 270144 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=163484 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169080 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=166598 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=169908 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=160748 MemoryPeaks { standard: 100352, interleaved: 65536, total: 165888, exchange_rows: 0, maximum_standard_allocation: 76800 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 400 }) COARSE_DROP old=144299 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=128184 MemoryPeaks { standard: 52600, interleaved: 204800, total: 237568, exchange_rows: 0, maximum_standard_allocation: 37376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 272 }) COARSE_REPLACED old=144299 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=137139 MemoryPeaks { standard: 60376, interleaved: 143872, total: 183808, exchange_rows: 0, maximum_standard_allocation: 34272 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 272 }) COARSE_CAP 137720 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=154856 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=123411 MemoryPeaks { standard: 80896, interleaved: 65536, total: 146432, exchange_rows: 0, maximum_standard_allocation: 40960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 160 }) COARSE_CAP 137720 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=146348 MemoryPeaks { standard: 53208, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=130332 MemoryPeaks { standard: 52600, interleaved: 204800, total: 237568, exchange_rows: 0, maximum_standard_allocation: 37376 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 272 }) COARSE_REPLACED old=146348 MemoryPeaks { standard: 60376, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 36288 } winner=139262 MemoryPeaks { standard: 60376, interleaved: 143872, total: 183808, exchange_rows: 0, maximum_standard_allocation: 34272 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 272 }) COARSE_CAP 139868 MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=157004 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=124124 MemoryPeaks { standard: 92600, interleaved: 65536, total: 149504, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 144 }) COARSE_CAP 139868 MemoryPeaks { standard: 51480, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=249809 MemoryPeaks { standard: 111648, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=249809 MemoryPeaks { standard: 111648, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387503 MemoryPeaks { standard: 275328, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387503 MemoryPeaks { standard: 275328, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521880 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=521880 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=482854 MemoryPeaks { standard: 246952, interleaved: 284672, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=534570 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=494275 MemoryPeaks { standard: 302248, interleaved: 229376, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=361687 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=337862 MemoryPeaks { standard: 235584, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=361687 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=337862 MemoryPeaks { standard: 235584, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=231264 MemoryPeaks { standard: 130048, interleaved: 98304, total: 228352, exchange_rows: 0, maximum_standard_allocation: 95040 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=374377 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=349283 MemoryPeaks { standard: 235584, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=7282714 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7250722 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384571 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352579 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=972441 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940449 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=919650 MemoryPeaks { standard: 575464, interleaved: 425984, total: 938048, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1111383 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079391 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7299850 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7267858 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401707 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369715 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=989577 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957585 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128519 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096527 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7282714 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7250722 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384571 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352579 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=972441 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940449 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1111383 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079391 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7299850 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7267858 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401707 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369715 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=989577 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957585 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128519 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096527 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_REPLACED old=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=803680 MemoryPeaks { standard: 450688, interleaved: 499712, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=775144 MemoryPeaks { standard: 443520, interleaved: 499712, total: 930560, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=818912 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=818912 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=7282714 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7250722 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384571 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352579 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=972441 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940449 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=919650 MemoryPeaks { standard: 579808, interleaved: 425984, total: 938048, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1111383 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079391 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7299850 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7267858 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401707 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369715 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=989577 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957585 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128519 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096527 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7282714 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7250722 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384571 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352579 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=972441 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940449 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1111383 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079391 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7299850 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7267858 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401707 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369715 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=989577 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957585 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128519 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955098 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096527 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411543 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_REPLACED old=875552 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=803680 MemoryPeaks { standard: 450688, interleaved: 499712, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=843560 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=775144 MemoryPeaks { standard: 443520, interleaved: 499712, total: 930560, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=892688 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=818912 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=860696 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=818912 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=229251 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=195905 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=229251 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=195905 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=431831 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=431831 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=447063 MemoryPeaks { standard: 228928, interleaved: 229376, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=7283004 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7251012 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384716 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352724 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=972584 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940592 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=919940 MemoryPeaks { standard: 575464, interleaved: 425984, total: 938048, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1111526 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079534 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7300140 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7268148 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401852 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369860 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=989720 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957728 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128662 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096670 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7283004 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7251012 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7384716 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352724 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=972584 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940592 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1111526 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079534 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7300140 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7268148 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401852 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369860 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=989720 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957728 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128662 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 490472, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096670 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_REPLACED old=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=803968 MemoryPeaks { standard: 450688, interleaved: 499712, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=775432 MemoryPeaks { standard: 443520, interleaved: 499712, total: 930560, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=819200 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=819200 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=201842 MemoryPeaks { standard: 94184, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=201842 MemoryPeaks { standard: 94184, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=222727 MemoryPeaks { standard: 103968, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=222727 MemoryPeaks { standard: 103968, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=194880 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=190099 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=152675 MemoryPeaks { standard: 76288, interleaved: 130048, total: 169984, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=190099 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=152675 MemoryPeaks { standard: 69120, interleaved: 194560, total: 227328, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=210126 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=191689 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=210126 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=191689 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=206921 MemoryPeaks { standard: 113664, interleaved: 98304, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4522822 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4498894 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4590839 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4566911 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=732647 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708719 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535512 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4511584 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4603529 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4579601 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=745337 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721409 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4522822 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4498894 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4590839 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4566911 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640702 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=640702 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=732647 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708719 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535512 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511584 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4603529 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4579601 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=653392 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=653392 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=745337 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=721409 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=664630 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620405 MemoryPeaks { standard: 438720, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=640702 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620405 MemoryPeaks { standard: 438720, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=677320 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=631826 MemoryPeaks { standard: 438720, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=653392 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=631826 MemoryPeaks { standard: 438720, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=7283004 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7251012 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7384716 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7352724 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=972584 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940592 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=919940 MemoryPeaks { standard: 579808, interleaved: 425984, total: 938048, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1111526 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079534 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7300140 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7268148 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7401852 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7369860 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=989720 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957728 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128662 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096670 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7283004 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7251012 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7384716 MemoryPeaks { standard: 13024192, interleaved: 508928, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7352724 MemoryPeaks { standard: 13017024, interleaved: 508928, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=972584 MemoryPeaks { standard: 859072, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=940592 MemoryPeaks { standard: 851904, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1111526 MemoryPeaks { standard: 1126912, interleaved: 508928, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1079534 MemoryPeaks { standard: 1119744, interleaved: 508928, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=7300140 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13030272, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7268148 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13106048, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7401852 MemoryPeaks { standard: 13024192, interleaved: 425984, total: 13024192, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=7369860 MemoryPeaks { standard: 13099968, interleaved: 425984, total: 13099968, exchange_rows: 0, maximum_standard_allocation: 12177216 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=989720 MemoryPeaks { standard: 859072, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=957728 MemoryPeaks { standard: 934848, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 419904 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1128662 MemoryPeaks { standard: 1126912, interleaved: 425984, total: 1272960, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=955388 MemoryPeaks { standard: 494816, interleaved: 425984, total: 853056, exchange_rows: 0, maximum_standard_allocation: 404352 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1096670 MemoryPeaks { standard: 1202688, interleaved: 425984, total: 1348736, exchange_rows: 0, maximum_standard_allocation: 676512 } winner=411888 MemoryPeaks { standard: 1200384, interleaved: 12288, total: 1200384, exchange_rows: 0, maximum_standard_allocation: 617472 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 384 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=875840 MemoryPeaks { standard: 453568, interleaved: 508928, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=803968 MemoryPeaks { standard: 450688, interleaved: 499712, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=843848 MemoryPeaks { standard: 446400, interleaved: 508928, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=775432 MemoryPeaks { standard: 443520, interleaved: 499712, total: 930560, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=892976 MemoryPeaks { standard: 453568, interleaved: 425984, total: 865152, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=819200 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=860984 MemoryPeaks { standard: 529344, interleaved: 425984, total: 940928, exchange_rows: 0, maximum_standard_allocation: 417600 } winner=819200 MemoryPeaks { standard: 450688, interleaved: 425984, total: 864000, exchange_rows: 0, maximum_standard_allocation: 418176 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681555 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728512 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522418 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587794 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=510764 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698691 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745648 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4522822 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4498894 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=495081 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4590839 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=539554 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4566911 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=732647 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708719 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535512 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511584 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4603529 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4579601 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=745337 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721409 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4522822 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4498894 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=539554 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=604930 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=529804 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=477945 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=431831 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=477945 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=431831 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=4590839 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=4566911 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640702 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=640702 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=732647 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708719 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535512 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511584 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4603529 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=224979 MemoryPeaks { standard: 145728, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=201842 MemoryPeaks { standard: 98528, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=495081 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=447063 MemoryPeaks { standard: 228928, interleaved: 229376, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=224979 MemoryPeaks { standard: 145728, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=201842 MemoryPeaks { standard: 98528, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=4579601 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=263501 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=653392 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=653392 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=745337 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721409 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=664630 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620405 MemoryPeaks { standard: 434376, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=640702 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620405 MemoryPeaks { standard: 434376, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=677320 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=631826 MemoryPeaks { standard: 434376, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=653392 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=631826 MemoryPeaks { standard: 434376, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=220882 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=208271 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196590 MemoryPeaks { standard: 134016, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=208271 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196590 MemoryPeaks { standard: 126848, interleaved: 208896, total: 274432, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=220961 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208011 MemoryPeaks { standard: 134016, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4523112 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4499184 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4590984 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4567056 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1104999 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=732792 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708864 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535802 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1117000 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4511874 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4603674 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197132 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213566 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4579746 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=745482 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=721554 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1122135 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4523112 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=4499184 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4590984 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4567056 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1134136 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207235 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214268 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640992 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=640992 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=732792 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708864 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535802 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511874 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4603674 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4579746 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=653682 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=653682 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=745482 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721554 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=230702 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230702 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=664920 MemoryPeaks { standard: 490472, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620693 MemoryPeaks { standard: 434376, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=640992 MemoryPeaks { standard: 483304, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620693 MemoryPeaks { standard: 434376, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=677610 MemoryPeaks { standard: 490472, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=632114 MemoryPeaks { standard: 434376, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=653682 MemoryPeaks { standard: 544744, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=632114 MemoryPeaks { standard: 434376, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=207235 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=194880 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=190099 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=152675 MemoryPeaks { standard: 76288, interleaved: 130048, total: 169984, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=190099 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=152675 MemoryPeaks { standard: 69120, interleaved: 194560, total: 227328, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387503 MemoryPeaks { standard: 270984, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387503 MemoryPeaks { standard: 270984, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401462 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2290802 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322363 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393248 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303492 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335053 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374377 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=405938 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=361687 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=337862 MemoryPeaks { standard: 231240, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=361687 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=337862 MemoryPeaks { standard: 231240, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=374377 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=349283 MemoryPeaks { standard: 231240, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294464 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294464 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=279142 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249579 MemoryPeaks { standard: 128128, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=279142 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249579 MemoryPeaks { standard: 120960, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=264811 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=4523112 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4499184 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4590984 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4567056 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=732792 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708864 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535802 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511874 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4603674 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4579746 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=745482 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721554 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4523112 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4499184 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4590984 MemoryPeaks { standard: 7845376, interleaved: 356352, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4567056 MemoryPeaks { standard: 7838208, interleaved: 356352, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640992 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=640992 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=732792 MemoryPeaks { standard: 567040, interleaved: 356352, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=708864 MemoryPeaks { standard: 559872, interleaved: 356352, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=4535802 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7860352, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4511874 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7914624, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4603674 MemoryPeaks { standard: 7845376, interleaved: 294912, total: 7845376, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=4579746 MemoryPeaks { standard: 7899648, interleaved: 294912, total: 7899648, exchange_rows: 0, maximum_standard_allocation: 7278336 } winner=264054 MemoryPeaks { standard: 754976, interleaved: 7680, total: 754976, exchange_rows: 0, maximum_standard_allocation: 385920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 }) COARSE_DROP old=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=653682 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=653682 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=745482 MemoryPeaks { standard: 567040, interleaved: 294912, total: 861952, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=721554 MemoryPeaks { standard: 621312, interleaved: 294912, total: 916224, exchange_rows: 0, maximum_standard_allocation: 279936 } winner=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=664920 MemoryPeaks { standard: 494816, interleaved: 356352, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620693 MemoryPeaks { standard: 438720, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=640992 MemoryPeaks { standard: 487648, interleaved: 356352, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=620693 MemoryPeaks { standard: 438720, interleaved: 350208, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=677610 MemoryPeaks { standard: 494816, interleaved: 294912, total: 609280, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=632114 MemoryPeaks { standard: 438720, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=653682 MemoryPeaks { standard: 549088, interleaved: 294912, total: 636288, exchange_rows: 0, maximum_standard_allocation: 269568 } winner=632114 MemoryPeaks { standard: 438720, interleaved: 294912, total: 590080, exchange_rows: 0, maximum_standard_allocation: 278400 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154719 MemoryPeaks { standard: 97544, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154719 MemoryPeaks { standard: 97544, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294754 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294754 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=146962 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325520 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=141873 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=133484 MemoryPeaks { standard: 74440, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=141873 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=133484 MemoryPeaks { standard: 74440, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325520 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325520 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325520 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144905 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=293819 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=222727 MemoryPeaks { standard: 99624, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=222727 MemoryPeaks { standard: 99624, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=279431 MemoryPeaks { standard: 134016, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249868 MemoryPeaks { standard: 128128, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=279431 MemoryPeaks { standard: 126848, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249868 MemoryPeaks { standard: 120960, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=265100 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=241767 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=326202 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=293819 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=266410 MemoryPeaks { standard: 169376, interleaved: 186368, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=306509 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=277831 MemoryPeaks { standard: 224672, interleaved: 131072, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=135312 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141828 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141828 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148002 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=135312 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=126864 MemoryPeaks { standard: 103840, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_REPLACED old=148002 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138285 MemoryPeaks { standard: 159136, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1275836 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287801 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409294 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409294 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409294 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409294 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222091 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238165 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1292972 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1304937 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227262 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239227 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255301 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=210126 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=191689 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=210126 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=191689 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=227262 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=206921 MemoryPeaks { standard: 113664, interleaved: 98304, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=373343 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=342446 MemoryPeaks { standard: 185760, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=386033 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=353867 MemoryPeaks { standard: 241056, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=330312 MemoryPeaks { standard: 168448, interleaved: 163840, total: 327680, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=172974 MemoryPeaks { standard: 111616, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55680 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255393 MemoryPeaks { standard: 166144, interleaved: 219136, total: 326528, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_REPLACED old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255393 MemoryPeaks { standard: 158976, interleaved: 274432, total: 374656, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=156703 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163423 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163423 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163423 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163423 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154719 MemoryPeaks { standard: 101888, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154719 MemoryPeaks { standard: 101888, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=168678 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=693882 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701995 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146962 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706572 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714685 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154563 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159652 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=141873 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=133484 MemoryPeaks { standard: 78784, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=141873 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=133484 MemoryPeaks { standard: 78784, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=154563 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144905 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=214295 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=202326 MemoryPeaks { standard: 136608, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=226985 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=213747 MemoryPeaks { standard: 191904, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=250098 MemoryPeaks { standard: 111648, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=250098 MemoryPeaks { standard: 111648, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=231554 MemoryPeaks { standard: 130048, interleaved: 98304, total: 228352, exchange_rows: 0, maximum_standard_allocation: 95040 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=640284 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625200 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=145292 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=135312 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141828 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652974 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637890 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=157136 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148002 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154518 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=229540 MemoryPeaks { standard: 106976, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=196232 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=229540 MemoryPeaks { standard: 99808, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=196232 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=135312 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=126864 MemoryPeaks { standard: 99496, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=148002 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138285 MemoryPeaks { standard: 154792, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=799741 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=807974 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=156703 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164936 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=128848 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812431 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820664 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169393 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177626 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138370 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294464 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294464 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312230 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314239 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877221 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900454 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297407 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297407 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325793 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=205665 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894357 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917590 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296278 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314543 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=342929 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=279142 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249579 MemoryPeaks { standard: 128128, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=279142 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249579 MemoryPeaks { standard: 120960, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=296278 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=264811 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 133536, interleaved: 126976, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 194976, interleaved: 65536, total: 194976, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=134771 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=126290 MemoryPeaks { standard: 103840, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=147461 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137711 MemoryPeaks { standard: 159136, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 197024, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=227274 MemoryPeaks { standard: 258464, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 204192, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 258464, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=214584 MemoryPeaks { standard: 142752, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=202616 MemoryPeaks { standard: 136608, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=227274 MemoryPeaks { standard: 204192, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214037 MemoryPeaks { standard: 191904, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387792 MemoryPeaks { standard: 275328, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387792 MemoryPeaks { standard: 275328, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 275328, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=254662 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=239985 MemoryPeaks { standard: 167680, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=267352 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=251406 MemoryPeaks { standard: 167680, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 317920, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=361976 MemoryPeaks { standard: 263648, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=338151 MemoryPeaks { standard: 235584, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=361976 MemoryPeaks { standard: 256480, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=338151 MemoryPeaks { standard: 235584, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=374666 MemoryPeaks { standard: 263648, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=349572 MemoryPeaks { standard: 235584, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=988224 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=988224 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 511144, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 872616, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=988224 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=988224 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=160880 MemoryPeaks { standard: 71904, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=160880 MemoryPeaks { standard: 71904, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 511144, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 872616, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=223016 MemoryPeaks { standard: 103968, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=223016 MemoryPeaks { standard: 103968, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325809 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325809 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325809 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325809 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=147943 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=147943 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379026 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224486 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224486 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224486 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224486 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 263648, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=224979 MemoryPeaks { standard: 141384, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=224979 MemoryPeaks { standard: 141384, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=150673 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141327 MemoryPeaks { standard: 67648, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 263648, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=150673 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141327 MemoryPeaks { standard: 60480, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=238938 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=210416 MemoryPeaks { standard: 100064, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=192040 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=210416 MemoryPeaks { standard: 92896, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=192040 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=156559 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=432120 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=432120 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=447352 MemoryPeaks { standard: 228928, interleaved: 229376, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=207272 MemoryPeaks { standard: 113664, interleaved: 98304, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=294108 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=266699 MemoryPeaks { standard: 165032, interleaved: 186368, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=306798 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=278120 MemoryPeaks { standard: 220328, interleaved: 131072, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=123815 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471314 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503187 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371787 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412125 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199352 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488450 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520323 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357050 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=388923 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=388923 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=429261 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429261 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=357050 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=330312 MemoryPeaks { standard: 168448, interleaved: 163840, total: 327680, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=198602 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211292 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=198602 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=176251 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1172504 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=225268 MemoryPeaks { standard: 145728, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=211292 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=187672 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1188105 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=225268 MemoryPeaks { standard: 145728, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220848 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185194 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200795 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220961 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=233538 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 145728, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233538 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=208271 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196590 MemoryPeaks { standard: 129672, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=208271 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196590 MemoryPeaks { standard: 122504, interleaved: 208896, total: 274432, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=220961 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208011 MemoryPeaks { standard: 129672, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_REPLACED old=339914 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255393 MemoryPeaks { standard: 166144, interleaved: 219136, total: 326528, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_REPLACED old=339914 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255393 MemoryPeaks { standard: 158976, interleaved: 274432, total: 374656, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674451 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675148 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=162998 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146775 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150537 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150537 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691587 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692284 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=169760 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158246 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163911 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163911 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167673 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 194144, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=208560 MemoryPeaks { standard: 139872, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196880 MemoryPeaks { standard: 134016, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=208560 MemoryPeaks { standard: 132704, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196880 MemoryPeaks { standard: 126848, interleaved: 208896, total: 274432, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=221250 MemoryPeaks { standard: 139872, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208301 MemoryPeaks { standard: 134016, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409583 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409583 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409583 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409583 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=373632 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=342735 MemoryPeaks { standard: 181416, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=386322 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=354156 MemoryPeaks { standard: 236712, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148497 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148497 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3681816 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6524608, interleaved: 312320, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3728628 MemoryPeaks { standard: 6517440, interleaved: 395264, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 442048, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=522562 MemoryPeaks { standard: 434880, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=587938 MemoryPeaks { standard: 567808, interleaved: 312320, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=587938 MemoryPeaks { standard: 560640, interleaved: 395264, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=511052 MemoryPeaks { standard: 259304, interleaved: 321536, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6543744, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3698952 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6619520, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6524608, interleaved: 229376, total: 6524608, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=3745764 MemoryPeaks { standard: 6600384, interleaved: 229376, total: 6600384, exchange_rows: 0, maximum_standard_allocation: 6096960 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=495370 MemoryPeaks { standard: 315072, interleaved: 229376, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 442048, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=539698 MemoryPeaks { standard: 517824, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 567808, interleaved: 229376, total: 657024, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=605074 MemoryPeaks { standard: 643584, interleaved: 229376, total: 732800, exchange_rows: 0, maximum_standard_allocation: 338720 } winner=530092 MemoryPeaks { standard: 259304, interleaved: 229376, total: 459776, exchange_rows: 0, maximum_standard_allocation: 209664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=478234 MemoryPeaks { standard: 239296, interleaved: 312320, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=432120 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=478234 MemoryPeaks { standard: 232128, interleaved: 395264, total: 536960, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=432120 MemoryPeaks { standard: 228928, interleaved: 303104, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=495370 MemoryPeaks { standard: 239296, interleaved: 229376, total: 461184, exchange_rows: 0, maximum_standard_allocation: 217152 } winner=447352 MemoryPeaks { standard: 228928, interleaved: 229376, total: 451968, exchange_rows: 0, maximum_standard_allocation: 209088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 515488, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 876960, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 515488, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 876960, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=292859 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1540726 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562087 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254662 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276023 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145180 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553416 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574777 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267352 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288713 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=254662 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=239985 MemoryPeaks { standard: 163336, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=267352 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=251406 MemoryPeaks { standard: 163336, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224775 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224775 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224775 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224775 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 188704, interleaved: 126976, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=211602 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 250144, interleaved: 65536, total: 250144, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_REPLACED old=198912 MemoryPeaks { standard: 142752, interleaved: 126976, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=176616 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=211602 MemoryPeaks { standard: 204192, interleaved: 65536, total: 204192, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=188037 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 111648, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 104480, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=173373 MemoryPeaks { standard: 111616, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55680 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 199616, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=242035 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_REPLACED old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=330602 MemoryPeaks { standard: 168448, interleaved: 163840, total: 327680, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 168288, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 229728, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=254952 MemoryPeaks { standard: 175456, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=240274 MemoryPeaks { standard: 167680, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_REPLACED old=267642 MemoryPeaks { standard: 175456, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=251695 MemoryPeaks { standard: 167680, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1225032 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1194864 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214584 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227616 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237722 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207554 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=227274 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240306 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=214584 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=202616 MemoryPeaks { standard: 132264, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=227274 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214037 MemoryPeaks { standard: 187560, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_REPLACED old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255682 MemoryPeaks { standard: 166144, interleaved: 219136, total: 326528, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_REPLACED old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255682 MemoryPeaks { standard: 158976, interleaved: 274432, total: 374656, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_REPLACED old=157524 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=147516 MemoryPeaks { standard: 69280, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=157524 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=147516 MemoryPeaks { standard: 62112, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1039360, interleaved: 126976, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=639999 MemoryPeaks { standard: 1062912, interleaved: 126976, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=625060 MemoryPeaks { standard: 1016256, interleaved: 126976, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144726 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134771 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141432 MemoryPeaks { standard: 129192, interleaved: 126976, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1039360, interleaved: 65536, total: 1068032, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=652689 MemoryPeaks { standard: 1062912, interleaved: 65536, total: 1091584, exchange_rows: 0, maximum_standard_allocation: 958464 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=637750 MemoryPeaks { standard: 1016256, interleaved: 65536, total: 1016256, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=156570 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147461 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154122 MemoryPeaks { standard: 190632, interleaved: 65536, total: 190632, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=134771 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=126290 MemoryPeaks { standard: 99496, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=147461 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137711 MemoryPeaks { standard: 154792, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409583 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409583 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409583 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409583 MemoryPeaks { standard: 298400, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394528 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334192 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373632 MemoryPeaks { standard: 324000, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406446 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399696 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372605 MemoryPeaks { standard: 228768, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2407218 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346882 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386322 MemoryPeaks { standard: 385440, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=419136 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412386 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=373632 MemoryPeaks { standard: 208288, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=342735 MemoryPeaks { standard: 185760, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=386322 MemoryPeaks { standard: 269728, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=354156 MemoryPeaks { standard: 241056, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=158112 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137664 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=170802 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=149085 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=223016 MemoryPeaks { standard: 99624, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=223016 MemoryPeaks { standard: 99624, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 74880 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=242056 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=154781 MemoryPeaks { standard: 67752, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=154781 MemoryPeaks { standard: 67752, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=170013 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=161356 MemoryPeaks { standard: 67560, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=161356 MemoryPeaks { standard: 67560, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=988224 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=988224 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1276116 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 515488, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 876960, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072512 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2185600, interleaved: 181248, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1287936 MemoryPeaks { standard: 2178432, interleaved: 264192, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=222236 MemoryPeaks { standard: 163840, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=222236 MemoryPeaks { standard: 156672, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=6891504 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=988224 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=988224 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108266 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=238310 MemoryPeaks { standard: 194560, interleaved: 181248, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1088016 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=7085202 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=238310 MemoryPeaks { standard: 187392, interleaved: 264192, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2213632, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1293252 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2289408, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2185600, interleaved: 98304, total: 2185600, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904194 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 515488, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000914 MemoryPeaks { standard: 876960, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120956 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1305072 MemoryPeaks { standard: 2261376, interleaved: 98304, total: 2261376, exchange_rows: 0, maximum_standard_allocation: 2037888 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1100706 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=350232 MemoryPeaks { standard: 99488, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=227552 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=239372 MemoryPeaks { standard: 163840, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=239372 MemoryPeaks { standard: 239616, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255446 MemoryPeaks { standard: 194560, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=255446 MemoryPeaks { standard: 270336, interleaved: 98304, total: 321792, exchange_rows: 0, maximum_standard_allocation: 113216 } winner=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=210416 MemoryPeaks { standard: 96448, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=192040 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=210416 MemoryPeaks { standard: 89280, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=192040 MemoryPeaks { standard: 87808, interleaved: 172032, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=227552 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=207272 MemoryPeaks { standard: 113664, interleaved: 98304, total: 211968, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387792 MemoryPeaks { standard: 270984, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=387792 MemoryPeaks { standard: 270984, interleaved: 231424, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=401751 MemoryPeaks { standard: 270984, interleaved: 163840, total: 369664, exchange_rows: 0, maximum_standard_allocation: 138240 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521592 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_REPLACED old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157116 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_REPLACED old=521592 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=482565 MemoryPeaks { standard: 251296, interleaved: 284672, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=534282 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=493986 MemoryPeaks { standard: 306592, interleaved: 229376, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 103968, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2291072 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3931648, interleaved: 225280, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2322488 MemoryPeaks { standard: 3924480, interleaved: 286720, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 292480, interleaved: 225280, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=393392 MemoryPeaks { standard: 285312, interleaved: 286720, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157136 MemoryPeaks { standard: 96800, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3955328, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2303762 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 4009600, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3931648, interleaved: 163840, total: 3931648, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2335178 MemoryPeaks { standard: 3985920, interleaved: 163840, total: 3985920, exchange_rows: 0, maximum_standard_allocation: 3644160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=374666 MemoryPeaks { standard: 313576, interleaved: 163840, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 292480, interleaved: 163840, total: 451328, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=406082 MemoryPeaks { standard: 346752, interleaved: 163840, total: 505600, exchange_rows: 0, maximum_standard_allocation: 140160 } winner=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=361976 MemoryPeaks { standard: 259304, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=338151 MemoryPeaks { standard: 231240, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=361976 MemoryPeaks { standard: 252136, interleaved: 286720, total: 417792, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=338151 MemoryPeaks { standard: 231240, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_REPLACED old=374666 MemoryPeaks { standard: 259304, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 139776 } winner=349572 MemoryPeaks { standard: 231240, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 144768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209373 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165312 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170526 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=170526 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=183216 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=165312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=156096 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=178002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=167517 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=161356 MemoryPeaks { standard: 71904, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=161356 MemoryPeaks { standard: 71904, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=978780 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986425 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=184901 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184901 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=194999 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190333 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=180396 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=995916 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1003561 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163972 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=196876 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163972 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=202037 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163972 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=202037 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163972 MemoryPeaks { standard: 179616, interleaved: 65536, total: 179616, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=212135 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212135 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_REPLACED old=179740 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=154781 MemoryPeaks { standard: 72096, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_REPLACED old=179740 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=154781 MemoryPeaks { standard: 64928, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=196876 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=170013 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=867514 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=876059 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165593 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165593 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179039 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=179039 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155384 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=884650 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=893195 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174184 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174184 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=182729 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196175 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158273 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=157048 MemoryPeaks { standard: 71872, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146994 MemoryPeaks { standard: 69280, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_REPLACED old=157048 MemoryPeaks { standard: 64704, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146994 MemoryPeaks { standard: 62112, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=674736 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1096384, interleaved: 148480, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=675288 MemoryPeaks { standard: 1089216, interleaved: 231424, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=163520 MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 94144, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147204 MemoryPeaks { standard: 86976, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 100864, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793558 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795838 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150966 MemoryPeaks { standard: 93696, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 56608 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 131616, interleaved: 126976, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1126784, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=691872 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1202560, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1096384, interleaved: 65536, total: 1096384, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=692424 MemoryPeaks { standard: 1172160, interleaved: 65536, total: 1172160, exchange_rows: 0, maximum_standard_allocation: 1018944 } winner=170326 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158436 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168174 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158820 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=144402 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=810694 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164340 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812974 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168102 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160734 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 193056, interleaved: 65536, total: 193056, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=168324 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185310 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=151188 MemoryPeaks { standard: 69568, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141884 MemoryPeaks { standard: 67648, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=151188 MemoryPeaks { standard: 62400, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141884 MemoryPeaks { standard: 60480, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_REPLACED old=168324 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=157116 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189639 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181719 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=835200 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=815088 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165312 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170526 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847890 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827778 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=178002 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183216 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=165312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=156096 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=178002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=167517 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=762034 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=777088 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=158112 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=173166 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774724 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789778 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170802 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185856 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=158112 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137664 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=170802 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=149085 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=160880 MemoryPeaks { standard: 67560, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=160880 MemoryPeaks { standard: 67560, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 44928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=179920 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=793274 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1310656, interleaved: 148480, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=795699 MemoryPeaks { standard: 1303488, interleaved: 231424, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 101056, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158066 MemoryPeaks { standard: 93888, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 119296, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167804 MemoryPeaks { standard: 112128, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 67744 } winner=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1334144, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=810410 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1409920, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1310656, interleaved: 65536, total: 1310656, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=812835 MemoryPeaks { standard: 1386432, interleaved: 65536, total: 1386432, exchange_rows: 0, maximum_standard_allocation: 1219392 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=167809 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=175202 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184940 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=167809 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=156559 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278318 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278318 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278318 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278318 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=249225 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=249225 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=224043 MemoryPeaks { standard: 152992, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_REPLACED old=261915 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=235464 MemoryPeaks { standard: 208288, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=199859 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=199859 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=199859 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=199859 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=174533 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164308 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=187223 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=175729 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148497 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148497 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148647 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546984 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552648 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552648 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=124248 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559674 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565338 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131274 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136938 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=147943 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=147943 MemoryPeaks { standard: 109056, interleaved: 65536, total: 174592, exchange_rows: 0, maximum_standard_allocation: 69120 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=148093 MemoryPeaks { standard: 107520, interleaved: 32768, total: 140288, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 878080, interleaved: 94208, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=546694 MemoryPeaks { standard: 870912, interleaved: 155648, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 878080, interleaved: 94208, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=552503 MemoryPeaks { standard: 870912, interleaved: 155648, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=133660 MemoryPeaks { standard: 71872, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=123815 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } winner=118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 878080, interleaved: 32768, total: 879744, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=559384 MemoryPeaks { standard: 932352, interleaved: 32768, total: 934016, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 878080, interleaved: 32768, total: 878080, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=565193 MemoryPeaks { standard: 932352, interleaved: 32768, total: 932352, exchange_rows: 0, maximum_standard_allocation: 808704 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=130696 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=136505 MemoryPeaks { standard: 155648, interleaved: 32768, total: 188416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 160160, interleaved: 126976, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 221600, interleaved: 65536, total: 221600, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=164879 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155620 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=177569 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=167041 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=200199 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=200199 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=200199 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=200199 MemoryPeaks { standard: 216480, interleaved: 65536, total: 216480, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 200096, interleaved: 65536, total: 200096, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=867804 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1453504, interleaved: 148480, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=876204 MemoryPeaks { standard: 1446336, interleaved: 231424, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=165924 MemoryPeaks { standard: 105664, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=165924 MemoryPeaks { standard: 98496, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=179370 MemoryPeaks { standard: 131584, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=179370 MemoryPeaks { standard: 124416, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 75168 } winner=155904 MemoryPeaks { standard: 77824, interleaved: 139264, total: 162816, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 384 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1472384, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=884940 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1548160, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1453504, interleaved: 65536, total: 1453504, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=893340 MemoryPeaks { standard: 1529280, interleaved: 65536, total: 1529280, exchange_rows: 0, maximum_standard_allocation: 1353024 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174660 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=174660 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=183060 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=196506 MemoryPeaks { standard: 131584, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=196506 MemoryPeaks { standard: 207360, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=158850 MemoryPeaks { standard: 91648, interleaved: 65536, total: 155648, exchange_rows: 0, maximum_standard_allocation: 81408 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 384 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 146848, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 165280, interleaved: 126976, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 163232, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 156320, interleaved: 126976, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 226720, interleaved: 65536, total: 226720, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=174912 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164736 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 217760, interleaved: 65536, total: 217760, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=187602 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=176157 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=157636 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137149 MemoryPeaks { standard: 120224, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=170326 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148570 MemoryPeaks { standard: 175520, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=800024 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1318912, interleaved: 126976, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=808112 MemoryPeaks { standard: 1311744, interleaved: 188416, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 99624, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=157136 MemoryPeaks { standard: 92456, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 105856, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=165224 MemoryPeaks { standard: 98688, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49920 } winner=129402 MemoryPeaks { standard: 73408, interleaved: 111616, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1337600, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=812714 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1391872, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1318912, interleaved: 65536, total: 1318912, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=820802 MemoryPeaks { standard: 1373184, interleaved: 65536, total: 1373184, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=169826 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=177914 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=138924 MemoryPeaks { standard: 86016, interleaved: 65536, total: 151552, exchange_rows: 0, maximum_standard_allocation: 63360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 240 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=577131 MemoryPeaks { standard: 363936, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=249809 MemoryPeaks { standard: 107304, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=249809 MemoryPeaks { standard: 107304, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=268849 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224775 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224775 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224775 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224775 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=615623 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=615623 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=620976 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=620976 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135326 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=133507 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=133507 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137348 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=137348 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=131789 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=628313 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=628313 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=633666 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633666 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146197 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146197 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144485 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150038 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150038 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=146750 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3564024 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473520 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521880 MemoryPeaks { standard: 450976, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=582090 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571776 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 65696, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576714 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486210 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534570 MemoryPeaks { standard: 512416, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594780 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584466 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=584466 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=521880 MemoryPeaks { standard: 273824, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=482854 MemoryPeaks { standard: 251296, interleaved: 284672, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_REPLACED old=534570 MemoryPeaks { standard: 335264, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=494275 MemoryPeaks { standard: 306592, interleaved: 229376, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513481 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1528902 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244672 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272698 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1066533 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067853 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198912 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221832 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221832 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179604 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=272698 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530617 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546038 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246387 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261808 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289834 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1079223 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080543 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211602 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234522 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=246387 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=231264 MemoryPeaks { standard: 130048, interleaved: 98304, total: 228352, exchange_rows: 0, maximum_standard_allocation: 95040 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=198912 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=176616 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=211602 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=188037 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=229251 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=195905 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=229251 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=195905 MemoryPeaks { standard: 92160, interleaved: 227328, total: 260096, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=181274 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1254400, interleaved: 126976, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=761750 MemoryPeaks { standard: 1284352, interleaved: 126976, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1239088, interleaved: 126976, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=776949 MemoryPeaks { standard: 1269040, interleaved: 126976, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=157636 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=172835 MemoryPeaks { standard: 151976, interleaved: 126976, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1254400, interleaved: 65536, total: 1275392, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=774440 MemoryPeaks { standard: 1284352, interleaved: 65536, total: 1305344, exchange_rows: 0, maximum_standard_allocation: 1158144 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1239088, interleaved: 65536, total: 1239088, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=789639 MemoryPeaks { standard: 1269040, interleaved: 65536, total: 1269040, exchange_rows: 0, maximum_standard_allocation: 1141920 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170326 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=185525 MemoryPeaks { standard: 213416, interleaved: 65536, total: 213416, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=157636 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=137149 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=170326 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=148570 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_REPLACED old=141312 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=132264 MemoryPeaks { standard: 99496, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=154002 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143685 MemoryPeaks { standard: 154792, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=241746 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 81920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2071552, interleaved: 159744, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1224752 MemoryPeaks { standard: 2125824, interleaved: 159744, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 1978240, interleaved: 159744, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1194729 MemoryPeaks { standard: 2032512, interleaved: 159744, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=214295 MemoryPeaks { standard: 192680, interleaved: 159744, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 154624, interleaved: 159744, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=227472 MemoryPeaks { standard: 208896, interleaved: 159744, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2071552, interleaved: 98304, total: 2096128, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1237442 MemoryPeaks { standard: 2125824, interleaved: 98304, total: 2150400, exchange_rows: 0, maximum_standard_allocation: 1916928 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 1978240, interleaved: 98304, total: 1978240, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1207419 MemoryPeaks { standard: 2032512, interleaved: 98304, total: 2032512, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=226985 MemoryPeaks { standard: 254120, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 199848, interleaved: 98304, total: 252928, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=240162 MemoryPeaks { standard: 254120, interleaved: 98304, total: 307200, exchange_rows: 0, maximum_standard_allocation: 73728 } winner=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=214295 MemoryPeaks { standard: 138408, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=202326 MemoryPeaks { standard: 132264, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=226985 MemoryPeaks { standard: 199848, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=213747 MemoryPeaks { standard: 187560, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163972 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163972 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163972 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163972 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603612 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603612 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630744 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185969 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131712 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=148044 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616302 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643434 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=144402 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160734 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409294 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=409294 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409294 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=409294 MemoryPeaks { standard: 294056, interleaved: 163840, total: 386048, exchange_rows: 0, maximum_standard_allocation: 147456 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4135936, interleaved: 225280, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2394258 MemoryPeaks { standard: 4251648, interleaved: 225280, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 3949312, interleaved: 225280, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2334067 MemoryPeaks { standard: 4065024, interleaved: 225280, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=373343 MemoryPeaks { standard: 319656, interleaved: 225280, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 312736, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=406302 MemoryPeaks { standard: 428448, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 302080, interleaved: 225280, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=399552 MemoryPeaks { standard: 417792, interleaved: 225280, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=372315 MemoryPeaks { standard: 224424, interleaved: 196608, total: 403456, exchange_rows: 0, maximum_standard_allocation: 194560 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4135936, interleaved: 163840, total: 4152320, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2406948 MemoryPeaks { standard: 4251648, interleaved: 163840, total: 4268032, exchange_rows: 0, maximum_standard_allocation: 3833856 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 3949312, interleaved: 163840, total: 3949312, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2346757 MemoryPeaks { standard: 4065024, interleaved: 163840, total: 4065024, exchange_rows: 0, maximum_standard_allocation: 3654144 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=386033 MemoryPeaks { standard: 381096, interleaved: 163840, total: 479232, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 312736, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=418992 MemoryPeaks { standard: 428448, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 302080, interleaved: 163840, total: 465920, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=412242 MemoryPeaks { standard: 417792, interleaved: 163840, total: 581632, exchange_rows: 0, maximum_standard_allocation: 147456 } winner=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=373343 MemoryPeaks { standard: 203944, interleaved: 225280, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=342446 MemoryPeaks { standard: 181416, interleaved: 219136, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=386033 MemoryPeaks { standard: 265384, interleaved: 163840, total: 363520, exchange_rows: 0, maximum_standard_allocation: 159744 } winner=353867 MemoryPeaks { standard: 236712, interleaved: 163840, total: 340992, exchange_rows: 0, maximum_standard_allocation: 155904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=189240 MemoryPeaks { standard: 195752, interleaved: 65536, total: 195752, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1383424, interleaved: 126976, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=834917 MemoryPeaks { standard: 1417216, interleaved: 126976, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1321216, interleaved: 126976, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=814950 MemoryPeaks { standard: 1355008, interleaved: 126976, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164879 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=170238 MemoryPeaks { standard: 155816, interleaved: 126976, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1383424, interleaved: 65536, total: 1399808, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=847607 MemoryPeaks { standard: 1417216, interleaved: 65536, total: 1433600, exchange_rows: 0, maximum_standard_allocation: 1277952 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1321216, interleaved: 65536, total: 1321216, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=827640 MemoryPeaks { standard: 1355008, interleaved: 65536, total: 1355008, exchange_rows: 0, maximum_standard_allocation: 1218048 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=177569 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=182928 MemoryPeaks { standard: 217256, interleaved: 65536, total: 217256, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=164879 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155620 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=177569 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=167041 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325520 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325520 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325520 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325520 MemoryPeaks { standard: 261288, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809505 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764398 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=293819 MemoryPeaks { standard: 256168, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=241827 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339162 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313512 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300226 MemoryPeaks { standard: 191656, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822195 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777088 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306509 MemoryPeaks { standard: 317608, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351852 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 232616, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326202 MemoryPeaks { standard: 317608, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=293819 MemoryPeaks { standard: 171176, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=266410 MemoryPeaks { standard: 165032, interleaved: 186368, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=306509 MemoryPeaks { standard: 232616, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=277831 MemoryPeaks { standard: 220328, interleaved: 131072, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201302 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947035 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957572 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176037 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=137831 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188086 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=166817 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=959725 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155184 MemoryPeaks { standard: 101888, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155184 MemoryPeaks { standard: 101888, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970262 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=188727 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=200776 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=188727 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=172974 MemoryPeaks { standard: 111616, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55680 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278607 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278607 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278607 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163423 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=163423 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163423 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278607 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=163423 MemoryPeaks { standard: 175272, interleaved: 65536, total: 175272, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 256 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=147312 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147312 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=142368 MemoryPeaks { standard: 98528, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=134016 MemoryPeaks { standard: 78784, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=142368 MemoryPeaks { standard: 91360, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=134016 MemoryPeaks { standard: 78784, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145437 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=603326 MemoryPeaks { standard: 974848, interleaved: 126976, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=603326 MemoryPeaks { standard: 996480, interleaved: 126976, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=630603 MemoryPeaks { standard: 992704, interleaved: 126976, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=630603 MemoryPeaks { standard: 1014336, interleaved: 126976, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=185548 MemoryPeaks { standard: 150400, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 71232 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 }) COARSE_DROP old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=131146 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147623 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=147623 MemoryPeaks { standard: 127272, interleaved: 126976, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=616016 MemoryPeaks { standard: 974848, interleaved: 65536, total: 1005824, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=701080 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=616016 MemoryPeaks { standard: 996480, interleaved: 65536, total: 1027456, exchange_rows: 0, maximum_standard_allocation: 898560 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=727234 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 992704, interleaved: 65536, total: 992704, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=141312 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=643293 MemoryPeaks { standard: 1014336, interleaved: 65536, total: 1014336, exchange_rows: 0, maximum_standard_allocation: 913536 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=143836 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160313 MemoryPeaks { standard: 188712, interleaved: 65536, total: 188712, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=249514 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=249514 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=224332 MemoryPeaks { standard: 148648, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_REPLACED old=262204 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=235753 MemoryPeaks { standard: 203944, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=164118 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150612 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713770 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739924 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=733783 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735344 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168405 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=154002 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=152418 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176808 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=152418 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176808 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=162456 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=159168 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159168 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=156725 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_REPLACED old=141312 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=132264 MemoryPeaks { standard: 103840, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=750919 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752480 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=154002 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143685 MemoryPeaks { standard: 159136, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=169554 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169554 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=174491 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=734068 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1203520, interleaved: 148480, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=735484 MemoryPeaks { standard: 1196352, interleaved: 231424, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=168896 MemoryPeaks { standard: 61352, interleaved: 163840, total: 203776, exchange_rows: 0, maximum_standard_allocation: 50688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 }) COARSE_DROP old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 97600, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=152820 MemoryPeaks { standard: 90432, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 110080, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=159570 MemoryPeaks { standard: 102912, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 62176 } winner=157264 MemoryPeaks { standard: 68608, interleaved: 151552, total: 175104, exchange_rows: 0, maximum_standard_allocation: 58368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1230464, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=751204 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1306240, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1203520, interleaved: 65536, total: 1203520, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=752620 MemoryPeaks { standard: 1279296, interleaved: 65536, total: 1279296, exchange_rows: 0, maximum_standard_allocation: 1119168 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=169956 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=176706 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=176706 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=175030 MemoryPeaks { standard: 109568, interleaved: 65536, total: 175104, exchange_rows: 0, maximum_standard_allocation: 86016 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=202152 MemoryPeaks { standard: 98528, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=202152 MemoryPeaks { standard: 98528, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=195304 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=202152 MemoryPeaks { standard: 94184, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=202152 MemoryPeaks { standard: 94184, interleaved: 157696, total: 214016, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=221192 MemoryPeaks { standard: 148480, interleaved: 65536, total: 214016, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=190464 MemoryPeaks { standard: 78784, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=153156 MemoryPeaks { standard: 76288, interleaved: 130048, total: 169984, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=190464 MemoryPeaks { standard: 71616, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=153156 MemoryPeaks { standard: 69120, interleaved: 194560, total: 227328, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1105281 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1882048, interleaved: 148480, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1117137 MemoryPeaks { standard: 1874880, interleaved: 231424, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 136768, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197352 MemoryPeaks { standard: 129600, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 168448, interleaved: 148480, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=213786 MemoryPeaks { standard: 161280, interleaved: 231424, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1887104, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1122417 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1962880, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1882048, interleaved: 65536, total: 1882048, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1134273 MemoryPeaks { standard: 1957824, interleaved: 65536, total: 1957824, exchange_rows: 0, maximum_standard_allocation: 1753920 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=207600 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 136768, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=214488 MemoryPeaks { standard: 212544, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 168448, interleaved: 65536, total: 193664, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=230922 MemoryPeaks { standard: 244224, interleaved: 65536, total: 269440, exchange_rows: 0, maximum_standard_allocation: 97440 } winner=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=207600 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=195304 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 148776, interleaved: 126976, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 101544, interleaved: 122880, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 210216, interleaved: 65536, total: 210216, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 158888, interleaved: 65536, total: 162816, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_REPLACED old=140800 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131717 MemoryPeaks { standard: 99496, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=153490 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143138 MemoryPeaks { standard: 154792, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=190464 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=153156 MemoryPeaks { standard: 76288, interleaved: 130048, total: 169984, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=190464 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=153156 MemoryPeaks { standard: 69120, interleaved: 194560, total: 227328, exchange_rows: 0, maximum_standard_allocation: 65664 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294754 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=294754 MemoryPeaks { standard: 160768, interleaved: 249856, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=312520 MemoryPeaks { standard: 160768, interleaved: 163840, total: 320768, exchange_rows: 0, maximum_standard_allocation: 145920 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=314528 MemoryPeaks { standard: 164864, interleaved: 131072, total: 295936, exchange_rows: 0, maximum_standard_allocation: 104832 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1877496 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3274816, interleaved: 214016, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1900584 MemoryPeaks { standard: 3267648, interleaved: 296960, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=297551 MemoryPeaks { standard: 233536, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=297551 MemoryPeaks { standard: 226368, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 288256, interleaved: 214016, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=325937 MemoryPeaks { standard: 281088, interleaved: 296960, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=206016 MemoryPeaks { standard: 166144, interleaved: 212992, total: 327680, exchange_rows: 0, maximum_standard_allocation: 154368 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 64, output_columns: 384 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3300480, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1894632 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3376256, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3274816, interleaved: 131072, total: 3274816, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1917720 MemoryPeaks { standard: 3350592, interleaved: 131072, total: 3350592, exchange_rows: 0, maximum_standard_allocation: 3056832 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=296567 MemoryPeaks { standard: 231424, interleaved: 131072, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 233536, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=314687 MemoryPeaks { standard: 309312, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 288256, interleaved: 131072, total: 349056, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=343073 MemoryPeaks { standard: 364032, interleaved: 131072, total: 424832, exchange_rows: 0, maximum_standard_allocation: 169824 } winner=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=279431 MemoryPeaks { standard: 132160, interleaved: 214016, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249868 MemoryPeaks { standard: 128128, interleaved: 204800, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=279431 MemoryPeaks { standard: 124992, interleaved: 296960, total: 362496, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=249868 MemoryPeaks { standard: 120960, interleaved: 278528, total: 327680, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=296567 MemoryPeaks { standard: 155648, interleaved: 131072, total: 286720, exchange_rows: 0, maximum_standard_allocation: 116928 } winner=265100 MemoryPeaks { standard: 130048, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 114048 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224486 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=224486 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224486 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=224486 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1792000, interleaved: 126976, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1066251 MemoryPeaks { standard: 1837952, interleaved: 126976, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1731856, interleaved: 126976, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1067716 MemoryPeaks { standard: 1777808, interleaved: 126976, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=198602 MemoryPeaks { standard: 184360, interleaved: 126976, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=221667 MemoryPeaks { standard: 138408, interleaved: 126976, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221667 MemoryPeaks { standard: 184360, interleaved: 126976, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=179183 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1792000, interleaved: 65536, total: 1793792, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1078941 MemoryPeaks { standard: 1837952, interleaved: 65536, total: 1839744, exchange_rows: 0, maximum_standard_allocation: 1657344 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1731856, interleaved: 65536, total: 1731856, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1080406 MemoryPeaks { standard: 1777808, interleaved: 65536, total: 1777808, exchange_rows: 0, maximum_standard_allocation: 1598688 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=211292 MemoryPeaks { standard: 245800, interleaved: 65536, total: 245800, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 199848, interleaved: 65536, total: 201728, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=234357 MemoryPeaks { standard: 245800, interleaved: 65536, total: 247680, exchange_rows: 0, maximum_standard_allocation: 65536 } winner=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=198602 MemoryPeaks { standard: 138408, interleaved: 126976, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=176251 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=211292 MemoryPeaks { standard: 199848, interleaved: 65536, total: 199848, exchange_rows: 0, maximum_standard_allocation: 79872 } winner=187672 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=250098 MemoryPeaks { standard: 107304, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=250098 MemoryPeaks { standard: 107304, interleaved: 190464, total: 246784, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=269138 MemoryPeaks { standard: 148480, interleaved: 98304, total: 246784, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325809 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=325809 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325809 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=325809 MemoryPeaks { standard: 265632, interleaved: 131072, total: 320512, exchange_rows: 0, maximum_standard_allocation: 114688 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1513760 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2614144, interleaved: 181248, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1529036 MemoryPeaks { standard: 2606976, interleaved: 264192, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 177664, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=244816 MemoryPeaks { standard: 170496, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=272842 MemoryPeaks { standard: 231424, interleaved: 181248, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=272842 MemoryPeaks { standard: 224256, interleaved: 264192, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2628352, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1530896 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2704128, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2614144, interleaved: 98304, total: 2614144, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1546172 MemoryPeaks { standard: 2689920, interleaved: 98304, total: 2689920, exchange_rows: 0, maximum_standard_allocation: 2438784 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=246676 MemoryPeaks { standard: 215040, interleaved: 98304, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261952 MemoryPeaks { standard: 177664, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261952 MemoryPeaks { standard: 253440, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 84096 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 231424, interleaved: 98304, total: 273664, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=289978 MemoryPeaks { standard: 307200, interleaved: 98304, total: 349440, exchange_rows: 0, maximum_standard_allocation: 135488 } winner=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=246676 MemoryPeaks { standard: 139264, interleaved: 98304, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=231554 MemoryPeaks { standard: 130048, interleaved: 98304, total: 228352, exchange_rows: 0, maximum_standard_allocation: 95040 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3103744, interleaved: 192512, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1809780 MemoryPeaks { standard: 3188736, interleaved: 192512, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 2963776, interleaved: 192512, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1764528 MemoryPeaks { standard: 3048768, interleaved: 192512, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=294108 MemoryPeaks { standard: 260512, interleaved: 192512, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=242116 MemoryPeaks { standard: 176128, interleaved: 163840, total: 336896, exchange_rows: 0, maximum_standard_allocation: 162816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 275872, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=339306 MemoryPeaks { standard: 360864, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 228352, interleaved: 192512, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=313656 MemoryPeaks { standard: 313344, interleaved: 192512, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=300516 MemoryPeaks { standard: 196000, interleaved: 163840, total: 330752, exchange_rows: 0, maximum_standard_allocation: 155648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3103744, interleaved: 131072, total: 3124224, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1822470 MemoryPeaks { standard: 3188736, interleaved: 131072, total: 3209216, exchange_rows: 0, maximum_standard_allocation: 2875392 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 2963776, interleaved: 131072, total: 2963776, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1777218 MemoryPeaks { standard: 3048768, interleaved: 131072, total: 3048768, exchange_rows: 0, maximum_standard_allocation: 2740608 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=306798 MemoryPeaks { standard: 321952, interleaved: 131072, total: 382976, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 275872, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=351996 MemoryPeaks { standard: 360864, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 236960, interleaved: 131072, total: 359424, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=326346 MemoryPeaks { standard: 321952, interleaved: 131072, total: 444416, exchange_rows: 0, maximum_standard_allocation: 110592 } winner=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=294108 MemoryPeaks { standard: 175520, interleaved: 192512, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=266699 MemoryPeaks { standard: 169376, interleaved: 186368, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=306798 MemoryPeaks { standard: 236960, interleaved: 131072, total: 297984, exchange_rows: 0, maximum_standard_allocation: 119808 } winner=278120 MemoryPeaks { standard: 224672, interleaved: 131072, total: 291840, exchange_rows: 0, maximum_standard_allocation: 111360 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=229540 MemoryPeaks { standard: 102632, interleaved: 181248, total: 237568, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=196232 MemoryPeaks { standard: 99328, interleaved: 162816, total: 202752, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_REPLACED old=229540 MemoryPeaks { standard: 95464, interleaved: 264192, total: 313344, exchange_rows: 0, maximum_standard_allocation: 83520 } winner=196232 MemoryPeaks { standard: 92160, interleaved: 227328, total: 260096, exchange_rows: 0, maximum_standard_allocation: 87552 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 288 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=293149 MemoryPeaks { standard: 195272, interleaved: 98304, total: 271360, exchange_rows: 0, maximum_standard_allocation: 98304 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1541016 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2619904, interleaved: 159744, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1562232 MemoryPeaks { standard: 2612736, interleaved: 221184, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=254952 MemoryPeaks { standard: 163944, interleaved: 221184, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 193792, interleaved: 159744, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=276168 MemoryPeaks { standard: 186624, interleaved: 221184, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=145734 MemoryPeaks { standard: 138208, interleaved: 154112, total: 267840, exchange_rows: 0, maximum_standard_allocation: 128160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 48, output_columns: 240 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2624896, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1553706 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2679168, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2619904, interleaved: 98304, total: 2619904, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1574922 MemoryPeaks { standard: 2674176, interleaved: 98304, total: 2674176, exchange_rows: 0, maximum_standard_allocation: 2426112 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=267642 MemoryPeaks { standard: 225384, interleaved: 98304, total: 303104, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 193792, interleaved: 98304, total: 292096, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=288858 MemoryPeaks { standard: 248064, interleaved: 98304, total: 346368, exchange_rows: 0, maximum_standard_allocation: 93312 } winner=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=254952 MemoryPeaks { standard: 171112, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=240274 MemoryPeaks { standard: 163336, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=267642 MemoryPeaks { standard: 171112, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 89856 } winner=251695 MemoryPeaks { standard: 163336, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 100224 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278607 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278607 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278607 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278607 MemoryPeaks { standard: 232864, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468532 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455140 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249514 MemoryPeaks { standard: 226208, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=323008 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232967 MemoryPeaks { standard: 163232, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286396 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249826 MemoryPeaks { standard: 196000, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1481222 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467830 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=262204 MemoryPeaks { standard: 287648, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335698 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 220576, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=299086 MemoryPeaks { standard: 287648, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=249514 MemoryPeaks { standard: 159136, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=224332 MemoryPeaks { standard: 152992, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=262204 MemoryPeaks { standard: 220576, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=235753 MemoryPeaks { standard: 208288, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=576842 MemoryPeaks { standard: 359592, interleaved: 229376, total: 517120, exchange_rows: 0, maximum_standard_allocation: 221184 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6200320, interleaved: 290816, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3563763 MemoryPeaks { standard: 6377472, interleaved: 290816, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 5920384, interleaved: 290816, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3473404 MemoryPeaks { standard: 6097536, interleaved: 290816, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=521592 MemoryPeaks { standard: 446632, interleaved: 290816, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 465520, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=581947 MemoryPeaks { standard: 642672, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 449536, interleaved: 290816, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=571633 MemoryPeaks { standard: 626688, interleaved: 290816, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6200320, interleaved: 229376, total: 6208512, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3576453 MemoryPeaks { standard: 6377472, interleaved: 229376, total: 6385664, exchange_rows: 0, maximum_standard_allocation: 5750784 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 5920384, interleaved: 229376, total: 5920384, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=3486094 MemoryPeaks { standard: 6097536, interleaved: 229376, total: 6097536, exchange_rows: 0, maximum_standard_allocation: 5481216 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=534282 MemoryPeaks { standard: 508072, interleaved: 229376, total: 671744, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 465520, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=594637 MemoryPeaks { standard: 642672, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 228384 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=584323 MemoryPeaks { standard: 449536, interleaved: 229376, total: 678912, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=584323 MemoryPeaks { standard: 626688, interleaved: 229376, total: 856064, exchange_rows: 0, maximum_standard_allocation: 221184 } winner=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_REPLACED old=521592 MemoryPeaks { standard: 269480, interleaved: 290816, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=482565 MemoryPeaks { standard: 246952, interleaved: 284672, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_REPLACED old=534282 MemoryPeaks { standard: 330920, interleaved: 229376, total: 494592, exchange_rows: 0, maximum_standard_allocation: 219648 } winner=493986 MemoryPeaks { standard: 302248, interleaved: 229376, total: 472064, exchange_rows: 0, maximum_standard_allocation: 222720 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=379316 MemoryPeaks { standard: 181248, interleaved: 163840, total: 345088, exchange_rows: 0, maximum_standard_allocation: 134784 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2471604 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4346176, interleaved: 246784, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2503332 MemoryPeaks { standard: 4339008, interleaved: 329728, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=371931 MemoryPeaks { standard: 302656, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=371931 MemoryPeaks { standard: 295488, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 380416, interleaved: 246784, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=412269 MemoryPeaks { standard: 373248, interleaved: 329728, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=199872 MemoryPeaks { standard: 216064, interleaved: 221184, total: 411136, exchange_rows: 0, maximum_standard_allocation: 205824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 32, output_columns: 384 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4370048, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2488740 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4445824, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4346176, interleaved: 163840, total: 4346176, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=2520468 MemoryPeaks { standard: 4421952, interleaved: 163840, total: 4421952, exchange_rows: 0, maximum_standard_allocation: 4059072 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=357339 MemoryPeaks { standard: 247808, interleaved: 163840, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 302656, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=389067 MemoryPeaks { standard: 378432, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 380416, interleaved: 163840, total: 450944, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=429405 MemoryPeaks { standard: 456192, interleaved: 163840, total: 526720, exchange_rows: 0, maximum_standard_allocation: 225504 } winner=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=357339 MemoryPeaks { standard: 172032, interleaved: 163840, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=330602 MemoryPeaks { standard: 168448, interleaved: 163840, total: 327680, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=340203 MemoryPeaks { standard: 167872, interleaved: 246784, total: 335872, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255682 MemoryPeaks { standard: 166144, interleaved: 219136, total: 326528, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_REPLACED old=340203 MemoryPeaks { standard: 160704, interleaved: 329728, total: 411648, exchange_rows: 0, maximum_standard_allocation: 150336 } winner=255682 MemoryPeaks { standard: 158976, interleaved: 274432, total: 374656, exchange_rows: 0, maximum_standard_allocation: 152064 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155184 MemoryPeaks { standard: 97544, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155184 MemoryPeaks { standard: 97544, interleaved: 133120, total: 189440, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=169143 MemoryPeaks { standard: 123904, interleaved: 65536, total: 189440, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=694166 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1136128, interleaved: 126976, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=702134 MemoryPeaks { standard: 1128960, interleaved: 188416, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=147312 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1161344, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=706856 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1215616, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1136128, interleaved: 65536, total: 1136128, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=714824 MemoryPeaks { standard: 1190400, interleaved: 65536, total: 1190400, exchange_rows: 0, maximum_standard_allocation: 1048320 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=155058 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=160002 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=142368 MemoryPeaks { standard: 94184, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=134016 MemoryPeaks { standard: 74440, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=142368 MemoryPeaks { standard: 87016, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 49152 } winner=134016 MemoryPeaks { standard: 74440, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 44544 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=155058 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145437 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=225268 MemoryPeaks { standard: 141384, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=225268 MemoryPeaks { standard: 141384, interleaved: 165888, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=239227 MemoryPeaks { standard: 141384, interleaved: 98304, total: 238592, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1172784 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1974784, interleaved: 159744, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1188240 MemoryPeaks { standard: 1967616, interleaved: 221184, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 150016, interleaved: 159744, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=220992 MemoryPeaks { standard: 142848, interleaved: 221184, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 2002816, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1185474 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2057088, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 1974784, interleaved: 98304, total: 1974784, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1200930 MemoryPeaks { standard: 2029056, interleaved: 98304, total: 2029056, exchange_rows: 0, maximum_standard_allocation: 1827072 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=221250 MemoryPeaks { standard: 189800, interleaved: 98304, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 150016, interleaved: 98304, total: 246016, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=233682 MemoryPeaks { standard: 204288, interleaved: 98304, total: 300288, exchange_rows: 0, maximum_standard_allocation: 70272 } winner=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=208560 MemoryPeaks { standard: 135528, interleaved: 159744, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196880 MemoryPeaks { standard: 129672, interleaved: 153600, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=208560 MemoryPeaks { standard: 128360, interleaved: 221184, total: 286720, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=196880 MemoryPeaks { standard: 122504, interleaved: 208896, total: 274432, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=221250 MemoryPeaks { standard: 135528, interleaved: 98304, total: 232448, exchange_rows: 0, maximum_standard_allocation: 69888 } winner=208301 MemoryPeaks { standard: 129672, interleaved: 98304, total: 226304, exchange_rows: 0, maximum_standard_allocation: 77952 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278318 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=278318 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278318 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=278318 MemoryPeaks { standard: 228520, interleaved: 98304, total: 254976, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2501632, interleaved: 159744, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1468253 MemoryPeaks { standard: 2568704, interleaved: 159744, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2388880, interleaved: 159744, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1455006 MemoryPeaks { standard: 2455952, interleaved: 159744, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=249225 MemoryPeaks { standard: 221864, interleaved: 159744, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 254368, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=322864 MemoryPeaks { standard: 321440, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 185344, interleaved: 159744, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=232677 MemoryPeaks { standard: 158888, interleaved: 131072, total: 276480, exchange_rows: 0, maximum_standard_allocation: 135168 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=286252 MemoryPeaks { standard: 252416, interleaved: 159744, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=249537 MemoryPeaks { standard: 191656, interleaved: 131072, total: 261120, exchange_rows: 0, maximum_standard_allocation: 116736 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2501632, interleaved: 98304, total: 2510848, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1480943 MemoryPeaks { standard: 2568704, interleaved: 98304, total: 2577920, exchange_rows: 0, maximum_standard_allocation: 2316288 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2388880, interleaved: 98304, total: 2388880, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=1467696 MemoryPeaks { standard: 2455952, interleaved: 98304, total: 2455952, exchange_rows: 0, maximum_standard_allocation: 2207712 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=261915 MemoryPeaks { standard: 283304, interleaved: 98304, total: 315904, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 254368, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=335554 MemoryPeaks { standard: 321440, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 152256 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 216232, interleaved: 98304, total: 283648, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=298942 MemoryPeaks { standard: 283304, interleaved: 98304, total: 350720, exchange_rows: 0, maximum_standard_allocation: 89088 } winner=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=249225 MemoryPeaks { standard: 154792, interleaved: 159744, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=224043 MemoryPeaks { standard: 148648, interleaved: 153600, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=261915 MemoryPeaks { standard: 216232, interleaved: 98304, total: 248832, exchange_rows: 0, maximum_standard_allocation: 99840 } winner=235464 MemoryPeaks { standard: 203944, interleaved: 98304, total: 242688, exchange_rows: 0, maximum_standard_allocation: 89088 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=199859 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=199859 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=199859 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=199859 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932376 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=909895 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174533 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169096 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184452 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184452 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182255 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945066 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922585 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187223 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197142 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=174533 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164308 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=187223 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=175729 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 73640, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=155260 MemoryPeaks { standard: 67752, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=155260 MemoryPeaks { standard: 67752, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=170492 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 511144, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 872616, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12393472, interleaved: 520192, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7072280 MemoryPeaks { standard: 12754944, interleaved: 520192, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 11833600, interleaved: 520192, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6891417 MemoryPeaks { standard: 12195072, interleaved: 520192, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 480256, interleaved: 520192, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=987937 MemoryPeaks { standard: 841728, interleaved: 520192, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 923872, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1108124 MemoryPeaks { standard: 1285344, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 891904, interleaved: 520192, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1087874 MemoryPeaks { standard: 1253376, interleaved: 520192, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12393472, interleaved: 458752, total: 12409856, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=7084970 MemoryPeaks { standard: 12754944, interleaved: 458752, total: 12771328, exchange_rows: 0, maximum_standard_allocation: 11501568 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 11833600, interleaved: 458752, total: 11833600, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=6904107 MemoryPeaks { standard: 12195072, interleaved: 458752, total: 12195072, exchange_rows: 0, maximum_standard_allocation: 10962432 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 511144, interleaved: 458752, total: 922112, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1000627 MemoryPeaks { standard: 872616, interleaved: 458752, total: 1283584, exchange_rows: 0, maximum_standard_allocation: 439296 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 923872, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1120814 MemoryPeaks { standard: 1285344, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 456768 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 891904, interleaved: 458752, total: 1350656, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=1100564 MemoryPeaks { standard: 1253376, interleaved: 458752, total: 1712128, exchange_rows: 0, maximum_standard_allocation: 442368 } winner=349653 MemoryPeaks { standard: 95144, interleaved: 447488, total: 503808, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 1440, output_columns: 144 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=209752 MemoryPeaks { standard: 132096, interleaved: 65536, total: 197632, exchange_rows: 0, maximum_standard_allocation: 92160 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=979062 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1649920, interleaved: 148480, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=986562 MemoryPeaks { standard: 1642752, interleaved: 231424, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 129280, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=185184 MemoryPeaks { standard: 122112, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 148480, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=195282 MemoryPeaks { standard: 141312, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=190712 MemoryPeaks { standard: 77984, interleaved: 157696, total: 197632, exchange_rows: 0, maximum_standard_allocation: 59904 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1662464, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=996198 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1738240, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1649920, interleaved: 65536, total: 1649920, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=1003698 MemoryPeaks { standard: 1725696, interleaved: 65536, total: 1725696, exchange_rows: 0, maximum_standard_allocation: 1536768 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=197304 MemoryPeaks { standard: 198656, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 129280, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=202320 MemoryPeaks { standard: 205056, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 148480, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_DROP old=212418 MemoryPeaks { standard: 224256, interleaved: 65536, total: 264192, exchange_rows: 0, maximum_standard_allocation: 85376 } winner=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=180168 MemoryPeaks { standard: 78592, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=155260 MemoryPeaks { standard: 72096, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=180168 MemoryPeaks { standard: 71424, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 66816 } winner=155260 MemoryPeaks { standard: 64928, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_REPLACED old=197304 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=170492 MemoryPeaks { standard: 113664, interleaved: 65536, total: 179200, exchange_rows: 0, maximum_standard_allocation: 73728 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 288 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=201607 MemoryPeaks { standard: 140288, interleaved: 65536, total: 205824, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=947317 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1576960, interleaved: 126976, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=957709 MemoryPeaks { standard: 1569792, interleaved: 188416, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 107304, interleaved: 126976, total: 183296, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=176384 MemoryPeaks { standard: 100136, interleaved: 188416, total: 237568, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 125440, interleaved: 126976, total: 184832, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=138385 MemoryPeaks { standard: 86368, interleaved: 98304, total: 183232, exchange_rows: 0, maximum_standard_allocation: 76320 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 240 }) COARSE_DROP old=188288 MemoryPeaks { standard: 118272, interleaved: 188416, total: 239104, exchange_rows: 0, maximum_standard_allocation: 59904 } winner=167252 MemoryPeaks { standard: 83968, interleaved: 152064, total: 192000, exchange_rows: 0, maximum_standard_allocation: 72960 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1586432, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=960007 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1640704, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1576960, interleaved: 65536, total: 1576960, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=970399 MemoryPeaks { standard: 1631232, interleaved: 65536, total: 1631232, exchange_rows: 0, maximum_standard_allocation: 1457664 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=189074 MemoryPeaks { standard: 172032, interleaved: 65536, total: 237568, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 125440, interleaved: 65536, total: 184832, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=200978 MemoryPeaks { standard: 179712, interleaved: 65536, total: 239104, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=189074 MemoryPeaks { standard: 117760, interleaved: 65536, total: 183296, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=173373 MemoryPeaks { standard: 111616, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 55680 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1146880, interleaved: 126976, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=700795 MemoryPeaks { standard: 1173632, interleaved: 126976, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1156960, interleaved: 126976, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=727094 MemoryPeaks { standard: 1183712, interleaved: 126976, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=140800 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=163751 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=163751 MemoryPeaks { standard: 153120, interleaved: 126976, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=150073 MemoryPeaks { standard: 105888, interleaved: 122880, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1146880, interleaved: 65536, total: 1171712, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=713485 MemoryPeaks { standard: 1173632, interleaved: 65536, total: 1198464, exchange_rows: 0, maximum_standard_allocation: 1058304 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1156960, interleaved: 65536, total: 1156960, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=739784 MemoryPeaks { standard: 1183712, interleaved: 65536, total: 1183712, exchange_rows: 0, maximum_standard_allocation: 1065792 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=153490 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=176441 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=176441 MemoryPeaks { standard: 214560, interleaved: 65536, total: 214560, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=161917 MemoryPeaks { standard: 163232, interleaved: 65536, total: 163232, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_REPLACED old=140800 MemoryPeaks { standard: 126368, interleaved: 126976, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=131717 MemoryPeaks { standard: 103840, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=153490 MemoryPeaks { standard: 187808, interleaved: 65536, total: 187808, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=143138 MemoryPeaks { standard: 159136, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 55296 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=200199 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=200199 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=200199 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=200199 MemoryPeaks { standard: 212136, interleaved: 65536, total: 212136, exchange_rows: 0, maximum_standard_allocation: 67584 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 192 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=134048 MemoryPeaks { standard: 73640, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=134048 MemoryPeaks { standard: 66472, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 60584, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 67752, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1555456, interleaved: 126976, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=932658 MemoryPeaks { standard: 1594368, interleaved: 126976, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1485472, interleaved: 126976, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=910032 MemoryPeaks { standard: 1524384, interleaved: 126976, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=174912 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=169572 MemoryPeaks { standard: 142504, interleaved: 98304, total: 197632, exchange_rows: 0, maximum_standard_allocation: 90112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 256 }) COARSE_DROP old=184686 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 }) COARSE_DROP old=184686 MemoryPeaks { standard: 160936, interleaved: 126976, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=182672 MemoryPeaks { standard: 158888, interleaved: 98304, total: 195584, exchange_rows: 0, maximum_standard_allocation: 77824 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 112, output_columns: 256 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1555456, interleaved: 65536, total: 1565696, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=945348 MemoryPeaks { standard: 1594368, interleaved: 65536, total: 1604608, exchange_rows: 0, maximum_standard_allocation: 1437696 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1485472, interleaved: 65536, total: 1485472, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=922722 MemoryPeaks { standard: 1524384, interleaved: 65536, total: 1524384, exchange_rows: 0, maximum_standard_allocation: 1370304 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=187602 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_DROP old=197376 MemoryPeaks { standard: 222376, interleaved: 65536, total: 222376, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 192 }) COARSE_REPLACED old=174912 MemoryPeaks { standard: 122024, interleaved: 126976, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=164736 MemoryPeaks { standard: 115880, interleaved: 120832, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 192 }) COARSE_REPLACED old=187602 MemoryPeaks { standard: 183464, interleaved: 65536, total: 183464, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=176157 MemoryPeaks { standard: 171176, interleaved: 65536, total: 177152, exchange_rows: 0, maximum_standard_allocation: 66816 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 192 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 996352, interleaved: 126976, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=615908 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 996352, interleaved: 126976, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=621116 MemoryPeaks { standard: 989184, interleaved: 188416, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=135900 MemoryPeaks { standard: 86656, interleaved: 120832, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 288 }) COARSE_DROP old=134048 MemoryPeaks { standard: 77984, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=134048 MemoryPeaks { standard: 70816, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 64928, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 83968, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=137744 MemoryPeaks { standard: 76800, interleaved: 188416, total: 221184, exchange_rows: 0, maximum_standard_allocation: 39936 } winner=132346 MemoryPeaks { standard: 72096, interleaved: 126976, total: 166912, exchange_rows: 0, maximum_standard_allocation: 47520 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 240 }) COARSE_DROP old=628598 MemoryPeaks { standard: 996352, interleaved: 65536, total: 1026560, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=628598 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1080832, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 996352, interleaved: 65536, total: 996352, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=633806 MemoryPeaks { standard: 1050624, interleaved: 65536, total: 1050624, exchange_rows: 0, maximum_standard_allocation: 918528 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=146738 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=146738 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=145042 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 240 }) COARSE_DROP old=150434 MemoryPeaks { standard: 101376, interleaved: 65536, total: 166912, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_DROP old=150434 MemoryPeaks { standard: 155648, interleaved: 65536, total: 221184, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=147324 MemoryPeaks { standard: 95232, interleaved: 65536, total: 160768, exchange_rows: 0, maximum_standard_allocation: 76032 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 288 }) COARSE_REPLACED old=157048 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=133660 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_REPLACED old=157048 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=133660 MemoryPeaks { standard: 60360, interleaved: 203776, total: 236544, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=121147 MemoryPeaks { standard: 99624, interleaved: 32768, total: 130048, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 128 }) COARSE_REPLACED old=150673 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=150673 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=120741 MemoryPeaks { standard: 99464, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_DROP old=145889 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141327 MemoryPeaks { standard: 67648, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=145889 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141327 MemoryPeaks { standard: 60480, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=163025 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=120741 MemoryPeaks { standard: 103808, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_REPLACED old=145889 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=157524 MemoryPeaks { standard: 67528, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=134214 MemoryPeaks { standard: 67528, interleaved: 134656, total: 174592, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_REPLACED old=145889 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=157524 MemoryPeaks { standard: 60360, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=134214 MemoryPeaks { standard: 60360, interleaved: 203776, total: 236544, exchange_rows: 0, maximum_standard_allocation: 41760 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 240 }) COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=120741 MemoryPeaks { standard: 99464, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_DROP old=131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=121724 MemoryPeaks { standard: 99624, interleaved: 32768, total: 130048, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 128 }) COARSE_CAP 141110 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_DROP old=158246 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=120741 MemoryPeaks { standard: 103808, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_CAP 141110 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_CAP 118006 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_CAP 118006 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_REPLACED old=151188 MemoryPeaks { standard: 65224, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=151188 MemoryPeaks { standard: 58056, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=121318 MemoryPeaks { standard: 99464, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_CAP 130696 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } COARSE_REPLACED old=146436 MemoryPeaks { standard: 63496, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_REPLACED old=146436 MemoryPeaks { standard: 56328, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 }) COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=121318 MemoryPeaks { standard: 99464, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_CAP 118006 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_CAP 141684 MemoryPeaks { standard: 54600, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_CAP 118584 MemoryPeaks { standard: 62952, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=146436 MemoryPeaks { standard: 67840, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141884 MemoryPeaks { standard: 67648, interleaved: 139264, total: 179200, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=146436 MemoryPeaks { standard: 60672, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } winner=141884 MemoryPeaks { standard: 60480, interleaved: 212992, total: 245760, exchange_rows: 0, maximum_standard_allocation: 57024 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 288 }) COARSE_DROP old=163572 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=121318 MemoryPeaks { standard: 103808, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_CAP 118006 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_CAP 141684 MemoryPeaks { standard: 58944, interleaved: 231424, total: 264192, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_DROP old=158820 MemoryPeaks { standard: 122880, interleaved: 65536, total: 188416, exchange_rows: 0, maximum_standard_allocation: 82944 } winner=121318 MemoryPeaks { standard: 103808, interleaved: 65536, total: 142336, exchange_rows: 0, maximum_standard_allocation: 67840 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 128 }) COARSE_CAP 118584 MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_CAP 118584 MemoryPeaks { standard: 67296, interleaved: 155648, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_CAP 131274 MemoryPeaks { standard: 101376, interleaved: 32768, total: 134144, exchange_rows: 0, maximum_standard_allocation: 61440 } COARSE_DROP old=131146 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=112496 MemoryPeaks { standard: 97184, interleaved: 97792, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 }) COARSE_DROP old=143836 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=119165 MemoryPeaks { standard: 129440, interleaved: 65536, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 112 }) COARSE_DROP old=131146 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=112496 MemoryPeaks { standard: 92840, interleaved: 97792, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 }) COARSE_DROP old=143836 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=124112 MemoryPeaks { standard: 134312, interleaved: 65536, total: 158720, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 128 }) COARSE_CAP 141110 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_CAP 118584 MemoryPeaks { standard: 74464, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 } COARSE_DROP old=131712 MemoryPeaks { standard: 105640, interleaved: 126976, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=113076 MemoryPeaks { standard: 92840, interleaved: 97792, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 }) COARSE_DROP old=144402 MemoryPeaks { standard: 167080, interleaved: 65536, total: 167080, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=124686 MemoryPeaks { standard: 134312, interleaved: 65536, total: 158720, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 128 }) COARSE_CAP 141684 MemoryPeaks { standard: 61768, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 50112 } COARSE_DROP old=131712 MemoryPeaks { standard: 109984, interleaved: 126976, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=113076 MemoryPeaks { standard: 97184, interleaved: 97792, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 }) COARSE_DROP old=144402 MemoryPeaks { standard: 171424, interleaved: 65536, total: 171424, exchange_rows: 0, maximum_standard_allocation: 61440 } winner=119745 MemoryPeaks { standard: 129440, interleaved: 65536, total: 154112, exchange_rows: 0, maximum_standard_allocation: 49152 } Some(Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 112 }) OLD_FRAGMENT position=0 retained=0 coarse=Some((ProgramCycles { total: 137720, exchange: 38434 }, MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 0, maximum_standard_allocation: 32768 })) detailed=Some((ProgramCycles { total: 194134, exchange: 45840 }, MemoryPeaks { standard: 58648, interleaved: 148480, total: 188416, exchange_rows: 5456, maximum_standard_allocation: 32768 })) CAT position=0 cost=229708 peak=MemoryPeaks { standard: 73848, interleaved: 93184, total: 165888, exchange_rows: 13664, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(64) }, AxisTiling { axis: FromEnd(2), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=255926 peak=MemoryPeaks { standard: 57464, interleaved: 98304, total: 154624, exchange_rows: 16976, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1458, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(2), partitions: 9, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=236716 peak=MemoryPeaks { standard: 73848, interleaved: 69632, total: 142336, exchange_rows: 13664, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(48) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=214524 peak=MemoryPeaks { standard: 104312, interleaved: 75776, total: 132096, exchange_rows: 9524, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 112 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 39, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(234) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(39) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=214288 peak=MemoryPeaks { standard: 131192, interleaved: 32768, total: 162816, exchange_rows: 9632, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 192, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(72) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=219638 peak=MemoryPeaks { standard: 106616, interleaved: 55296, total: 160768, exchange_rows: 10028, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1428, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 17, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(84) }, AxisTiling { axis: FromEnd(2), partitions: 7, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=322956 peak=MemoryPeaks { standard: 62464, interleaved: 108544, total: 148480, exchange_rows: 24284, maximum_standard_allocation: 46080 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1452, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 33, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(44) }, AxisTiling { axis: FromEnd(2), partitions: 11, padding_groups: 1, block_size: 112, padding_multiple: 112, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=263044 peak=MemoryPeaks { standard: 100472, interleaved: 32768, total: 132096, exchange_rows: 15320, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 192, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 35, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=206804 peak=MemoryPeaks { standard: 68408, interleaved: 112128, total: 152064, exchange_rows: 8876, maximum_standard_allocation: 49920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 208 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1386, replicas: 21, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(231) }, AxisTiling { axis: FromEnd(1), partitions: 11, padding_groups: 1, block_size: 112, padding_multiple: 112, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(21) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=193952 peak=MemoryPeaks { standard: 57600, interleaved: 173056, total: 205824, exchange_rows: 6680, maximum_standard_allocation: 48000 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 112, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1386, replicas: 18, axes: [AxisTiling { axis: FromEnd(2), partitions: 7, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(198) }, AxisTiling { axis: FromEnd(1), partitions: 11, padding_groups: 1, block_size: 112, padding_multiple: 112, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(18) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=497323 peak=MemoryPeaks { standard: 77824, interleaved: 124416, total: 147968, exchange_rows: 43868, maximum_standard_allocation: 61824 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 49, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 80, padding_multiple: 80, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=197242 peak=MemoryPeaks { standard: 60376, interleaved: 143872, total: 183808, exchange_rows: 6212, maximum_standard_allocation: 34272 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 272 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1408, replicas: 16, axes: [AxisTiling { axis: FromEnd(2), partitions: 11, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(128) }, AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=231328 peak=MemoryPeaks { standard: 59896, interleaved: 137216, total: 177152, exchange_rows: 8084, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 272 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 16, axes: [AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(96) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=509488 peak=MemoryPeaks { standard: 83576, interleaved: 65536, total: 147968, exchange_rows: 43868, maximum_standard_allocation: 61824 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 80, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 49, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 80, padding_multiple: 80, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=230638 peak=MemoryPeaks { standard: 49504, interleaved: 171520, total: 195072, exchange_rows: 7796, maximum_standard_allocation: 32928 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(96) }, AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=209298 peak=MemoryPeaks { standard: 58688, interleaved: 159744, total: 183296, exchange_rows: 5816, maximum_standard_allocation: 41216 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 13, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(108) }, AxisTiling { axis: FromEnd(1), partitions: 9, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=216868 peak=MemoryPeaks { standard: 35328, interleaved: 269312, total: 285696, exchange_rows: 6536, maximum_standard_allocation: 26496 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 336 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1456, replicas: 13, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(91) }, AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(13) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=242168 peak=MemoryPeaks { standard: 57592, interleaved: 155648, total: 195584, exchange_rows: 7400, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 320 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 14, axes: [AxisTiling { axis: FromEnd(2), partitions: 17, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(84) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(14) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=243258 peak=MemoryPeaks { standard: 43904, interleaved: 161792, total: 185344, exchange_rows: 7112, maximum_standard_allocation: 26240 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 336 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 13, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(78) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(13) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=220578 peak=MemoryPeaks { standard: 33152, interleaved: 315392, total: 331776, exchange_rows: 5780, maximum_standard_allocation: 23680 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 20, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(72) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=292952 peak=MemoryPeaks { standard: 37176, interleaved: 240128, total: 263680, exchange_rows: 5780, maximum_standard_allocation: 19968 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 432 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1450, replicas: 10, axes: [AxisTiling { axis: FromEnd(2), partitions: 29, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(50) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=254360 peak=MemoryPeaks { standard: 30008, interleaved: 447488, total: 463872, exchange_rows: 4628, maximum_standard_allocation: 19968 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 432 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1450, replicas: 10, axes: [AxisTiling { axis: FromEnd(2), partitions: 29, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(50) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=594854 peak=MemoryPeaks { standard: 652800, interleaved: 11776, total: 652800, exchange_rows: 47360, maximum_standard_allocation: 327680 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 10, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(144) }, AxisTiling { axis: FromEnd(2), partitions: 72, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=597287 peak=MemoryPeaks { standard: 652800, interleaved: 0, total: 652800, exchange_rows: 47360, maximum_standard_allocation: 327680 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 10, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(144) }, AxisTiling { axis: FromEnd(2), partitions: 72, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=309284 peak=MemoryPeaks { standard: 35736, interleaved: 263168, total: 286720, exchange_rows: 5816, maximum_standard_allocation: 17664 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 480 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 32, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(45) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=283742 peak=MemoryPeaks { standard: 28224, interleaved: 555008, total: 571392, exchange_rows: 4052, maximum_standard_allocation: 18816 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 544 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 8, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(40) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=304316 peak=MemoryPeaks { standard: 27648, interleaved: 631808, total: 648192, exchange_rows: 3764, maximum_standard_allocation: 18432 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 624 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1435, replicas: 7, axes: [AxisTiling { axis: FromEnd(2), partitions: 41, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(35) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=307434 peak=MemoryPeaks { standard: 26744, interleaved: 659456, total: 675840, exchange_rows: 3440, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 544 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(2), partitions: 46, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(32) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=394346 peak=MemoryPeaks { standard: 33088, interleaved: 378368, total: 401920, exchange_rows: 5240, maximum_standard_allocation: 17280 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=329834 peak=MemoryPeaks { standard: 25920, interleaved: 723968, total: 740352, exchange_rows: 3296, maximum_standard_allocation: 17280 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=369656 peak=MemoryPeaks { standard: 24440, interleaved: 862208, total: 878592, exchange_rows: 3044, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1464, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 61, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(24) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=425016 peak=MemoryPeaks { standard: 23288, interleaved: 1028096, total: 1044480, exchange_rows: 2936, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 864 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=201990 peak=MemoryPeaks { standard: 73848, interleaved: 93184, total: 165888, exchange_rows: 10568, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(64) }, AxisTiling { axis: FromEnd(2), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=219568 peak=MemoryPeaks { standard: 57464, interleaved: 98304, total: 154624, exchange_rows: 12908, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1458, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(2), partitions: 9, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=202742 peak=MemoryPeaks { standard: 73848, interleaved: 83968, total: 156672, exchange_rows: 10928, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 15, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(96) }, AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=190134 peak=MemoryPeaks { standard: 106616, interleaved: 57344, total: 162816, exchange_rows: 7688, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(72) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=205902 peak=MemoryPeaks { standard: 98424, interleaved: 65536, total: 162816, exchange_rows: 10496, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 128, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(72) }, AxisTiling { axis: FromEnd(2), partitions: 9, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=207278 peak=MemoryPeaks { standard: 104312, interleaved: 75776, total: 132096, exchange_rows: 8984, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 112 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 39, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(234) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(39) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=222770 peak=MemoryPeaks { standard: 57464, interleaved: 75776, total: 132096, exchange_rows: 11864, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 35, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=195210 peak=MemoryPeaks { standard: 131192, interleaved: 32768, total: 162816, exchange_rows: 7688, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 192, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(72) }, AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=221735 peak=MemoryPeaks { standard: 89720, interleaved: 65536, total: 154112, exchange_rows: 11720, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1456, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 26, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(56) }, AxisTiling { axis: FromEnd(2), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=200912 peak=MemoryPeaks { standard: 106616, interleaved: 55296, total: 160768, exchange_rows: 8120, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1428, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 17, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(84) }, AxisTiling { axis: FromEnd(2), partitions: 7, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=219519 peak=MemoryPeaks { standard: 107640, interleaved: 32768, total: 139264, exchange_rows: 10784, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 176, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1456, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 26, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(56) }, AxisTiling { axis: FromEnd(2), partitions: 7, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=267524 peak=MemoryPeaks { standard: 77944, interleaved: 65536, total: 142336, exchange_rows: 18128, maximum_standard_allocation: 50688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(48) }, AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=215904 peak=MemoryPeaks { standard: 49504, interleaved: 171520, total: 195072, exchange_rows: 6212, maximum_standard_allocation: 32928 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(96) }, AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=206196 peak=MemoryPeaks { standard: 58688, interleaved: 159744, total: 183296, exchange_rows: 5888, maximum_standard_allocation: 41216 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 13, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(108) }, AxisTiling { axis: FromEnd(1), partitions: 9, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=202062 peak=MemoryPeaks { standard: 35328, interleaved: 269312, total: 285696, exchange_rows: 4988, maximum_standard_allocation: 26496 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 336 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1456, replicas: 13, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(91) }, AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(13) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=229362 peak=MemoryPeaks { standard: 57592, interleaved: 155648, total: 195584, exchange_rows: 6032, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 320 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 14, axes: [AxisTiling { axis: FromEnd(2), partitions: 17, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(84) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(14) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=224038 peak=MemoryPeaks { standard: 40440, interleaved: 161792, total: 185344, exchange_rows: 5384, maximum_standard_allocation: 23520 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 336 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1404, replicas: 13, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(78) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(13) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=261910 peak=MemoryPeaks { standard: 37408, interleaved: 217088, total: 240640, exchange_rows: 5060, maximum_standard_allocation: 21600 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 480 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=232256 peak=MemoryPeaks { standard: 32408, interleaved: 386048, total: 402432, exchange_rows: 4340, maximum_standard_allocation: 19840 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 368 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 12, axes: [AxisTiling { axis: FromEnd(2), partitions: 24, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(60) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=284626 peak=MemoryPeaks { standard: 37176, interleaved: 240128, total: 263680, exchange_rows: 5024, maximum_standard_allocation: 19968 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 432 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1450, replicas: 10, axes: [AxisTiling { axis: FromEnd(2), partitions: 29, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(50) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=246034 peak=MemoryPeaks { standard: 30008, interleaved: 447488, total: 463872, exchange_rows: 3872, maximum_standard_allocation: 19968 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 432 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1450, replicas: 10, axes: [AxisTiling { axis: FromEnd(2), partitions: 29, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(50) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=488719 peak=MemoryPeaks { standard: 652800, interleaved: 11776, total: 652800, exchange_rows: 34724, maximum_standard_allocation: 327680 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 10, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(144) }, AxisTiling { axis: FromEnd(2), partitions: 72, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=491152 peak=MemoryPeaks { standard: 652800, interleaved: 0, total: 652800, exchange_rows: 34724, maximum_standard_allocation: 327680 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 10, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(144) }, AxisTiling { axis: FromEnd(2), partitions: 72, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=301102 peak=MemoryPeaks { standard: 35736, interleaved: 263168, total: 286720, exchange_rows: 5132, maximum_standard_allocation: 17664 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 480 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 32, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(45) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=276496 peak=MemoryPeaks { standard: 28224, interleaved: 555008, total: 571392, exchange_rows: 3476, maximum_standard_allocation: 18816 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 544 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 8, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(40) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=352054 peak=MemoryPeaks { standard: 34816, interleaved: 332288, total: 355840, exchange_rows: 4772, maximum_standard_allocation: 18432 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 624 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1435, replicas: 7, axes: [AxisTiling { axis: FromEnd(2), partitions: 41, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(35) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=301316 peak=MemoryPeaks { standard: 26744, interleaved: 659456, total: 675840, exchange_rows: 3008, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 544 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(2), partitions: 46, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(32) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=388156 peak=MemoryPeaks { standard: 33088, interleaved: 378368, total: 401920, exchange_rows: 4844, maximum_standard_allocation: 17280 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=323644 peak=MemoryPeaks { standard: 25920, interleaved: 723968, total: 740352, exchange_rows: 2900, maximum_standard_allocation: 17280 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=364690 peak=MemoryPeaks { standard: 24440, interleaved: 862208, total: 878592, exchange_rows: 2792, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 720 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1464, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 61, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(24) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=420594 peak=MemoryPeaks { standard: 23288, interleaved: 1028096, total: 1044480, exchange_rows: 2792, maximum_standard_allocation: 16384 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 864 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=0 cost=391188 peak=MemoryPeaks { standard: 77824, interleaved: 124416, total: 147968, exchange_rows: 30872, maximum_standard_allocation: 61824 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 80, output_columns: 368 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(1), partitions: 49, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 80, padding_multiple: 80, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1472, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1472, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } OLD_FRAGMENT position=2 retained=0 coarse=Some((ProgramCycles { total: 118006, exchange: 32904 }, MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 0, maximum_standard_allocation: 32768 })) detailed=Some((ProgramCycles { total: 171190, exchange: 40104 }, MemoryPeaks { standard: 70120, interleaved: 94208, total: 134144, exchange_rows: 5960, maximum_standard_allocation: 32768 })) CAT position=2 cost=224402 peak=MemoryPeaks { standard: 64928, interleaved: 97792, total: 154112, exchange_rows: 19316, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=177922 peak=MemoryPeaks { standard: 103968, interleaved: 73728, total: 130048, exchange_rows: 8084, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(243) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=201853 peak=MemoryPeaks { standard: 97696, interleaved: 66048, total: 155136, exchange_rows: 13376, maximum_standard_allocation: 81920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 80 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 10, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=218600 peak=MemoryPeaks { standard: 81312, interleaved: 83968, total: 156672, exchange_rows: 17480, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(360) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=231071 peak=MemoryPeaks { standard: 97184, interleaved: 65536, total: 154112, exchange_rows: 19316, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=187168 peak=MemoryPeaks { standard: 130464, interleaved: 60416, total: 182272, exchange_rows: 8948, maximum_standard_allocation: 114688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 48 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 16, axes: [AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(240) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=221912 peak=MemoryPeaks { standard: 118176, interleaved: 32768, total: 142336, exchange_rows: 14060, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 192, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(184) }, AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=306322 peak=MemoryPeaks { standard: 85408, interleaved: 65536, total: 142336, exchange_rows: 30584, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=206016 peak=MemoryPeaks { standard: 77824, interleaved: 65536, total: 142336, exchange_rows: 10784, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1350, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 5, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(270) }, AxisTiling { axis: FromEnd(1), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=187524 peak=MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 6464, maximum_standard_allocation: 50112 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(120) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=194792 peak=MemoryPeaks { standard: 71968, interleaved: 124928, total: 164864, exchange_rows: 8516, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1380, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(115) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=195600 peak=MemoryPeaks { standard: 66208, interleaved: 143360, total: 183296, exchange_rows: 5240, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(92) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=178312 peak=MemoryPeaks { standard: 60032, interleaved: 235520, total: 268288, exchange_rows: 4916, maximum_standard_allocation: 41472 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(100) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=203590 peak=MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 7580, maximum_standard_allocation: 50688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(102) }, AxisTiling { axis: FromEnd(1), partitions: 34, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=182604 peak=MemoryPeaks { standard: 47616, interleaved: 286720, total: 303104, exchange_rows: 5744, maximum_standard_allocation: 44544 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(90) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=185966 peak=MemoryPeaks { standard: 43008, interleaved: 278528, total: 294912, exchange_rows: 5204, maximum_standard_allocation: 39936 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(81) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=221676 peak=MemoryPeaks { standard: 47104, interleaved: 167936, total: 191488, exchange_rows: 5924, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=194700 peak=MemoryPeaks { standard: 39936, interleaved: 303104, total: 319488, exchange_rows: 4952, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=234116 peak=MemoryPeaks { standard: 45472, interleaved: 192512, total: 216064, exchange_rows: 4808, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=204068 peak=MemoryPeaks { standard: 38304, interleaved: 352256, total: 368640, exchange_rows: 3872, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=261312 peak=MemoryPeaks { standard: 45120, interleaved: 217088, total: 240640, exchange_rows: 6392, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=224736 peak=MemoryPeaks { standard: 37952, interleaved: 401408, total: 417792, exchange_rows: 5096, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=261834 peak=MemoryPeaks { standard: 40800, interleaved: 217088, total: 240640, exchange_rows: 4880, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=223530 peak=MemoryPeaks { standard: 33632, interleaved: 401408, total: 417792, exchange_rows: 3368, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=529182 peak=MemoryPeaks { standard: 747808, interleaved: 7680, total: 755488, exchange_rows: 65104, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=530769 peak=MemoryPeaks { standard: 755488, interleaved: 0, total: 755488, exchange_rows: 65104, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=533705 peak=MemoryPeaks { standard: 754976, interleaved: 0, total: 754976, exchange_rows: 65612, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=363242 peak=MemoryPeaks { standard: 32448, interleaved: 843776, total: 860160, exchange_rows: 4448, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 57, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(25) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=409828 peak=MemoryPeaks { standard: 30976, interleaved: 991232, total: 1007616, exchange_rows: 4016, maximum_standard_allocation: 28160 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1407, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 67, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(21) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=432112 peak=MemoryPeaks { standard: 29472, interleaved: 1064960, total: 1081344, exchange_rows: 3980, maximum_standard_allocation: 24320 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 224, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 224, padding_multiple: 224, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=452020 peak=MemoryPeaks { standard: 29312, interleaved: 1138688, total: 1155072, exchange_rows: 3764, maximum_standard_allocation: 19584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 81, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(18) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=247975 peak=MemoryPeaks { standard: 105888, interleaved: 32768, total: 130048, exchange_rows: 19784, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1458, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 9, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(162) }, AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=196178 peak=MemoryPeaks { standard: 64928, interleaved: 97792, total: 154112, exchange_rows: 15860, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=176512 peak=MemoryPeaks { standard: 103968, interleaved: 73728, total: 130048, exchange_rows: 8120, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(243) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=180613 peak=MemoryPeaks { standard: 97696, interleaved: 66048, total: 155136, exchange_rows: 11000, maximum_standard_allocation: 81920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 80 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 10, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=193868 peak=MemoryPeaks { standard: 81312, interleaved: 83968, total: 156672, exchange_rows: 14492, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 96, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 4, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(360) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=202847 peak=MemoryPeaks { standard: 97184, interleaved: 65536, total: 154112, exchange_rows: 15860, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=174830 peak=MemoryPeaks { standard: 130464, interleaved: 60416, total: 182272, exchange_rows: 7508, maximum_standard_allocation: 114688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 48 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 16, axes: [AxisTiling { axis: FromEnd(1), partitions: 6, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(240) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=196016 peak=MemoryPeaks { standard: 118176, interleaved: 32768, total: 142336, exchange_rows: 10928, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 192, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1472, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(184) }, AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=263837 peak=MemoryPeaks { standard: 85408, interleaved: 65536, total: 142336, exchange_rows: 25364, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=204958 peak=MemoryPeaks { standard: 77824, interleaved: 65536, total: 142336, exchange_rows: 10856, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1350, replicas: 6, axes: [AxisTiling { axis: FromEnd(2), partitions: 5, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(270) }, AxisTiling { axis: FromEnd(1), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=186802 peak=MemoryPeaks { standard: 66112, interleaved: 148480, total: 188416, exchange_rows: 6536, maximum_standard_allocation: 50112 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(120) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=193998 peak=MemoryPeaks { standard: 71968, interleaved: 124928, total: 164864, exchange_rows: 8552, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1380, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(115) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=194734 peak=MemoryPeaks { standard: 66208, interleaved: 143360, total: 183296, exchange_rows: 5240, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(92) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=177590 peak=MemoryPeaks { standard: 60032, interleaved: 235520, total: 268288, exchange_rows: 4988, maximum_standard_allocation: 41472 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(100) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=203108 peak=MemoryPeaks { standard: 60928, interleaved: 163840, total: 187392, exchange_rows: 7616, maximum_standard_allocation: 50688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(102) }, AxisTiling { axis: FromEnd(1), partitions: 34, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=182122 peak=MemoryPeaks { standard: 47616, interleaved: 286720, total: 303104, exchange_rows: 5780, maximum_standard_allocation: 44544 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(90) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=185484 peak=MemoryPeaks { standard: 43008, interleaved: 278528, total: 294912, exchange_rows: 5240, maximum_standard_allocation: 39936 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(81) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=221194 peak=MemoryPeaks { standard: 47104, interleaved: 167936, total: 191488, exchange_rows: 5960, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=194218 peak=MemoryPeaks { standard: 39936, interleaved: 303104, total: 319488, exchange_rows: 4988, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=233634 peak=MemoryPeaks { standard: 45472, interleaved: 192512, total: 216064, exchange_rows: 4844, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=203586 peak=MemoryPeaks { standard: 38304, interleaved: 352256, total: 368640, exchange_rows: 3908, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=255982 peak=MemoryPeaks { standard: 45120, interleaved: 217088, total: 240640, exchange_rows: 5780, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=219406 peak=MemoryPeaks { standard: 37952, interleaved: 401408, total: 417792, exchange_rows: 4484, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=261256 peak=MemoryPeaks { standard: 40800, interleaved: 217088, total: 240640, exchange_rows: 4880, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=222952 peak=MemoryPeaks { standard: 33632, interleaved: 401408, total: 417792, exchange_rows: 3368, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=528388 peak=MemoryPeaks { standard: 747808, interleaved: 7680, total: 755488, exchange_rows: 65140, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=529975 peak=MemoryPeaks { standard: 755488, interleaved: 0, total: 755488, exchange_rows: 65140, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=532911 peak=MemoryPeaks { standard: 754976, interleaved: 0, total: 754976, exchange_rows: 65648, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=359520 peak=MemoryPeaks { standard: 32448, interleaved: 843776, total: 860160, exchange_rows: 4232, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 57, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(25) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=407010 peak=MemoryPeaks { standard: 30976, interleaved: 991232, total: 1007616, exchange_rows: 3944, maximum_standard_allocation: 28160 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1407, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 67, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(21) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=429262 peak=MemoryPeaks { standard: 29472, interleaved: 1064960, total: 1081344, exchange_rows: 3908, maximum_standard_allocation: 24320 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 224, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 224, padding_multiple: 224, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=449138 peak=MemoryPeaks { standard: 29312, interleaved: 1138688, total: 1155072, exchange_rows: 3728, maximum_standard_allocation: 19584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 81, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(18) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=218587 peak=MemoryPeaks { standard: 105888, interleaved: 32768, total: 130048, exchange_rows: 16184, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1458, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 9, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(162) }, AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=227570 peak=MemoryPeaks { standard: 60584, interleaved: 97792, total: 154112, exchange_rows: 19712, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=181378 peak=MemoryPeaks { standard: 99624, interleaved: 73728, total: 130048, exchange_rows: 8516, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(243) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=205021 peak=MemoryPeaks { standard: 93352, interleaved: 66048, total: 155136, exchange_rows: 13772, maximum_standard_allocation: 81920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 80 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 10, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=198172 peak=MemoryPeaks { standard: 109736, interleaved: 65536, total: 171008, exchange_rows: 11684, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 256, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1428, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(204) }, AxisTiling { axis: FromEnd(2), partitions: 17, padding_groups: 1, block_size: 256, padding_multiple: 256, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=190856 peak=MemoryPeaks { standard: 126120, interleaved: 55808, total: 177664, exchange_rows: 10208, maximum_standard_allocation: 114688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 48 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 16, axes: [AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(288) }, AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=250480 peak=MemoryPeaks { standard: 97448, interleaved: 65536, total: 158720, exchange_rows: 19352, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=251497 peak=MemoryPeaks { standard: 85160, interleaved: 65536, total: 146432, exchange_rows: 23996, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1350, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(270) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=309202 peak=MemoryPeaks { standard: 81064, interleaved: 65536, total: 142336, exchange_rows: 30944, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=236078 peak=MemoryPeaks { standard: 132264, interleaved: 32768, total: 160768, exchange_rows: 13196, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 288, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 12, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(120) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=254881 peak=MemoryPeaks { standard: 121512, interleaved: 32768, total: 150016, exchange_rows: 16472, maximum_standard_allocation: 60928 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 272, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1456, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 13, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(112) }, AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 272, padding_multiple: 272, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=197672 peak=MemoryPeaks { standard: 67624, interleaved: 124928, total: 164864, exchange_rows: 8876, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1380, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(115) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=175728 peak=MemoryPeaks { standard: 54696, interleaved: 253952, total: 286720, exchange_rows: 4700, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(92) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=180040 peak=MemoryPeaks { standard: 55688, interleaved: 235520, total: 268288, exchange_rows: 5132, maximum_standard_allocation: 41472 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(100) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=182374 peak=MemoryPeaks { standard: 53760, interleaved: 262144, total: 278528, exchange_rows: 6644, maximum_standard_allocation: 50688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(102) }, AxisTiling { axis: FromEnd(1), partitions: 34, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=183756 peak=MemoryPeaks { standard: 47616, interleaved: 286720, total: 303104, exchange_rows: 5888, maximum_standard_allocation: 44544 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(90) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=187118 peak=MemoryPeaks { standard: 43008, interleaved: 278528, total: 294912, exchange_rows: 5348, maximum_standard_allocation: 39936 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(81) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=223116 peak=MemoryPeaks { standard: 47104, interleaved: 167936, total: 191488, exchange_rows: 6104, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=199896 peak=MemoryPeaks { standard: 36864, interleaved: 327680, total: 344064, exchange_rows: 4700, maximum_standard_allocation: 33792 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(69) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=205796 peak=MemoryPeaks { standard: 33960, interleaved: 352256, total: 368640, exchange_rows: 4088, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=263040 peak=MemoryPeaks { standard: 40776, interleaved: 217088, total: 240640, exchange_rows: 6608, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=226464 peak=MemoryPeaks { standard: 33608, interleaved: 401408, total: 417792, exchange_rows: 5312, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=262986 peak=MemoryPeaks { standard: 39424, interleaved: 217088, total: 240640, exchange_rows: 5024, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=224682 peak=MemoryPeaks { standard: 32256, interleaved: 401408, total: 417792, exchange_rows: 3512, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=300792 peak=MemoryPeaks { standard: 36736, interleaved: 272384, total: 295936, exchange_rows: 6176, maximum_standard_allocation: 26880 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=255576 peak=MemoryPeaks { standard: 29568, interleaved: 512000, total: 528384, exchange_rows: 4772, maximum_standard_allocation: 26880 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=530334 peak=MemoryPeaks { standard: 747808, interleaved: 7680, total: 755488, exchange_rows: 65248, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=531921 peak=MemoryPeaks { standard: 755488, interleaved: 0, total: 755488, exchange_rows: 65248, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=289190 peak=MemoryPeaks { standard: 29376, interleaved: 622592, total: 638976, exchange_rows: 4052, maximum_standard_allocation: 26112 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 256, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1462, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 43, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(34) }, AxisTiling { axis: FromEnd(1), partitions: 17, padding_groups: 1, block_size: 256, padding_multiple: 256, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=299900 peak=MemoryPeaks { standard: 29352, interleaved: 659456, total: 675840, exchange_rows: 3980, maximum_standard_allocation: 23040 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 272, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 46, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(32) }, AxisTiling { axis: FromEnd(1), partitions: 16, padding_groups: 1, block_size: 272, padding_multiple: 272, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=309924 peak=MemoryPeaks { standard: 29288, interleaved: 696320, total: 712704, exchange_rows: 3944, maximum_standard_allocation: 20160 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=432808 peak=MemoryPeaks { standard: 26880, interleaved: 1064960, total: 1081344, exchange_rows: 3836, maximum_standard_allocation: 24320 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 224, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 224, padding_multiple: 224, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=452676 peak=MemoryPeaks { standard: 24968, interleaved: 1138688, total: 1155072, exchange_rows: 3656, maximum_standard_allocation: 19584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 81, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(18) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 729, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 729, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Reject, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=199346 peak=MemoryPeaks { standard: 60584, interleaved: 97792, total: 154112, exchange_rows: 16256, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=179968 peak=MemoryPeaks { standard: 99624, interleaved: 73728, total: 130048, exchange_rows: 8552, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 128 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 9, axes: [AxisTiling { axis: FromEnd(2), partitions: 6, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(243) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(9) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=183781 peak=MemoryPeaks { standard: 93352, interleaved: 66048, total: 155136, exchange_rows: 11396, maximum_standard_allocation: 81920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 80 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1470, replicas: 10, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(210) }, AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(10) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=181210 peak=MemoryPeaks { standard: 109736, interleaved: 65536, total: 171008, exchange_rows: 9740, maximum_standard_allocation: 98304 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 256, output_columns: 64 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1428, replicas: 12, axes: [AxisTiling { axis: FromEnd(1), partitions: 7, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(204) }, AxisTiling { axis: FromEnd(2), partitions: 17, padding_groups: 1, block_size: 256, padding_multiple: 256, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(12) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=178518 peak=MemoryPeaks { standard: 126120, interleaved: 55808, total: 177664, exchange_rows: 8768, maximum_standard_allocation: 114688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 48 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 16, axes: [AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(288) }, AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(16) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=221965 peak=MemoryPeaks { standard: 97448, interleaved: 65536, total: 158720, exchange_rows: 15860, maximum_standard_allocation: 49152 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 144, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=223273 peak=MemoryPeaks { standard: 85160, interleaved: 65536, total: 146432, exchange_rows: 20540, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 128 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1350, replicas: 6, axes: [AxisTiling { axis: FromEnd(1), partitions: 5, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(270) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(6) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=266717 peak=MemoryPeaks { standard: 81064, interleaved: 65536, total: 142336, exchange_rows: 25724, maximum_standard_allocation: 67584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 96, output_columns: 192 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 4, axes: [AxisTiling { axis: FromEnd(1), partitions: 8, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(180) }, AxisTiling { axis: FromEnd(2), partitions: 45, padding_groups: 1, block_size: 96, padding_multiple: 96, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=211346 peak=MemoryPeaks { standard: 132264, interleaved: 32768, total: 160768, exchange_rows: 10208, maximum_standard_allocation: 65536 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 288, output_columns: 96 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1440, replicas: 8, axes: [AxisTiling { axis: FromEnd(1), partitions: 12, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(120) }, AxisTiling { axis: FromEnd(2), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(8) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=226657 peak=MemoryPeaks { standard: 121512, interleaved: 32768, total: 150016, exchange_rows: 13052, maximum_standard_allocation: 60928 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 272, output_columns: 112 } input=Layout { order: Amp(TransposedLeft), tiling: TensorTiling { tile_count: 1456, replicas: 7, axes: [AxisTiling { axis: FromEnd(1), partitions: 13, padding_groups: 1, block_size: 2, padding_multiple: 2, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(112) }, AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 272, padding_multiple: 272, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(7) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=196878 peak=MemoryPeaks { standard: 67624, interleaved: 124928, total: 164864, exchange_rows: 8912, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1380, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 12, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(115) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=174862 peak=MemoryPeaks { standard: 54696, interleaved: 253952, total: 286720, exchange_rows: 4700, maximum_standard_allocation: 32768 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(92) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=179318 peak=MemoryPeaks { standard: 55688, interleaved: 235520, total: 268288, exchange_rows: 5204, maximum_standard_allocation: 41472 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 288 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1400, replicas: 4, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(100) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(4) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=181892 peak=MemoryPeaks { standard: 53760, interleaved: 262144, total: 278528, exchange_rows: 6680, maximum_standard_allocation: 50688 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 128, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1428, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 14, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(102) }, AxisTiling { axis: FromEnd(1), partitions: 34, padding_groups: 1, block_size: 128, padding_multiple: 128, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=183274 peak=MemoryPeaks { standard: 47616, interleaved: 286720, total: 303104, exchange_rows: 5924, maximum_standard_allocation: 44544 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1440, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 16, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(90) }, AxisTiling { axis: FromEnd(1), partitions: 30, padding_groups: 1, block_size: 144, padding_multiple: 144, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=186636 peak=MemoryPeaks { standard: 43008, interleaved: 278528, total: 294912, exchange_rows: 5384, maximum_standard_allocation: 39936 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 18, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(81) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=222634 peak=MemoryPeaks { standard: 47104, interleaved: 167936, total: 191488, exchange_rows: 6140, maximum_standard_allocation: 36864 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 176, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1425, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 19, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(75) }, AxisTiling { axis: FromEnd(1), partitions: 25, padding_groups: 1, block_size: 176, padding_multiple: 176, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=199414 peak=MemoryPeaks { standard: 36864, interleaved: 327680, total: 344064, exchange_rows: 4736, maximum_standard_allocation: 33792 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 21, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(69) }, AxisTiling { axis: FromEnd(1), partitions: 23, padding_groups: 1, block_size: 192, padding_multiple: 192, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=205314 peak=MemoryPeaks { standard: 33960, interleaved: 352256, total: 368640, exchange_rows: 4124, maximum_standard_allocation: 30720 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1449, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 23, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(63) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=257710 peak=MemoryPeaks { standard: 40776, interleaved: 217088, total: 240640, exchange_rows: 5996, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=221134 peak=MemoryPeaks { standard: 33608, interleaved: 401408, total: 417792, exchange_rows: 4700, maximum_standard_allocation: 29376 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 384 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 3, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(3) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=262408 peak=MemoryPeaks { standard: 39424, interleaved: 217088, total: 240640, exchange_rows: 5024, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=224104 peak=MemoryPeaks { standard: 32256, interleaved: 401408, total: 417792, exchange_rows: 3512, maximum_standard_allocation: 29952 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 27, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(54) }, AxisTiling { axis: FromEnd(1), partitions: 27, padding_groups: 1, block_size: 160, padding_multiple: 160, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=295366 peak=MemoryPeaks { standard: 36736, interleaved: 272384, total: 295936, exchange_rows: 5564, maximum_standard_allocation: 26880 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=250150 peak=MemoryPeaks { standard: 29568, interleaved: 512000, total: 528384, exchange_rows: 4160, maximum_standard_allocation: 26880 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 208, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 35, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(42) }, AxisTiling { axis: FromEnd(1), partitions: 21, padding_groups: 1, block_size: 208, padding_multiple: 208, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=529540 peak=MemoryPeaks { standard: 747808, interleaved: 7680, total: 755488, exchange_rows: 65284, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=531127 peak=MemoryPeaks { standard: 755488, interleaved: 0, total: 755488, exchange_rows: 65284, maximum_standard_allocation: 385920 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Standard, inner_block: 16, output_columns: 240 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1345, replicas: 5, axes: [AxisTiling { axis: FromEnd(2), partitions: 1, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }, AxisTiling { axis: FromEnd(1), partitions: 269, padding_groups: 1, block_size: 16, padding_multiple: 16, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(5) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=284412 peak=MemoryPeaks { standard: 29376, interleaved: 622592, total: 638976, exchange_rows: 3656, maximum_standard_allocation: 26112 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 256, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1462, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 43, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(34) }, AxisTiling { axis: FromEnd(1), partitions: 17, padding_groups: 1, block_size: 256, padding_multiple: 256, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=295674 peak=MemoryPeaks { standard: 29352, interleaved: 659456, total: 675840, exchange_rows: 3620, maximum_standard_allocation: 23040 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 272, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1472, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 46, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(32) }, AxisTiling { axis: FromEnd(1), partitions: 16, padding_groups: 1, block_size: 272, padding_multiple: 272, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=306250 peak=MemoryPeaks { standard: 29288, interleaved: 696320, total: 712704, exchange_rows: 3620, maximum_standard_allocation: 20160 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 288, output_columns: 576 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1470, replicas: 2, axes: [AxisTiling { axis: FromEnd(2), partitions: 49, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(30) }, AxisTiling { axis: FromEnd(1), partitions: 15, padding_groups: 1, block_size: 288, padding_multiple: 288, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(2) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=429958 peak=MemoryPeaks { standard: 26880, interleaved: 1064960, total: 1081344, exchange_rows: 3764, maximum_standard_allocation: 24320 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 224, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1460, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 73, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(20) }, AxisTiling { axis: FromEnd(1), partitions: 20, padding_groups: 1, block_size: 224, padding_multiple: 224, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } CAT position=2 cost=449794 peak=MemoryPeaks { standard: 24968, interleaved: 1138688, total: 1155072, exchange_rows: 3620, maximum_standard_allocation: 19584 } kernel=Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 240, output_columns: 1152 } input=Layout { order: Amp(Left), tiling: TensorTiling { tile_count: 1458, replicas: 1, axes: [AxisTiling { axis: FromEnd(2), partitions: 81, padding_groups: 1, block_size: 1, padding_multiple: 1, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(18) }, AxisTiling { axis: FromEnd(1), partitions: 18, padding_groups: 1, block_size: 240, padding_multiple: 240, shard_padding_multiple: 1, padding: Zero, tile_stride: Some(1) }] }, memory_class: Ipu21Standard } output=Layout { order: RowMajor, tiling: TensorTiling { tile_count: 1469, replicas: 1, axes: [AxisTiling { axis: Linear, partitions: 1469, padding_groups: 1, block_size: 4, padding_multiple: 4, shard_padding_multiple: 1, padding: Zero, tile_stride: None }] }, memory_class: Ipu21Standard } RESULT normal estimated=376366 memory=MemoryPeaks { standard: 106616, interleaved: 253952, total: 293888, exchange_rows: 12388, maximum_standard_allocation: 98304 } RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } Some((ProgramCycles { total: 48984, exchange: 21336 }, MemoryUsage { standard: 0, interleaved: 0 }, 868)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 39096, exchange: 6744 }, MemoryUsage { standard: 0, interleaved: 0 }, 3280)) RESULT_OP normal Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 64 } Some((ProgramCycles { total: 49446, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7512, exchange: 6648 }, MemoryUsage { standard: 0, interleaved: 0 }, 472)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7736, exchange: 6872 }, MemoryUsage { standard: 0, interleaved: 0 }, 688)) RESULT_OP normal ReductionSum { partials: 6 } Some((ProgramCycles { total: 6354, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 31006, exchange: 1752 }, MemoryUsage { standard: 4608, interleaved: 0 }, 2380)) RESULT_OP normal Gelu Some((ProgramCycles { total: 11370, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 31476, exchange: 5016 }, MemoryUsage { standard: 17664, interleaved: 0 }, 1840)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 42360, exchange: 28248 }, MemoryUsage { standard: 0, interleaved: 0 }, 256)) RESULT_OP normal Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } Some((ProgramCycles { total: 74598, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7368, exchange: 6936 }, MemoryUsage { standard: 0, interleaved: 0 }, 904)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7512, exchange: 7080 }, MemoryUsage { standard: 0, interleaved: 0 }, 1552)) RESULT_OP normal ReductionSum { partials: 23 } Some((ProgramCycles { total: 9510, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP normal Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 2038, exchange: 888 }, MemoryUsage { standard: 0, interleaved: 0 }, 148)) LOW_COST normal Ok(ProgramCycles { total: 284320, exchange: 76880 }) PREFIX normal 0 delta=21336 exchange=21336 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } test planner::gemm::tests::trace_fp16_regression has been running for over 60 seconds PREFIX normal 1 delta=36006 exchange=6144 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX normal 2 delta=49446 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 64 } PREFIX normal 3 delta=8784 exchange=6648 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX normal 4 delta=224 exchange=224 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX normal 5 delta=6354 exchange=0 ReductionSum { partials: 6 } PREFIX normal 6 delta=12867 exchange=1752 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX normal 7 delta=12219 exchange=0 Gelu PREFIX normal 8 delta=15738 exchange=5016 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX normal 9 delta=27648 exchange=27648 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX normal 10 delta=74598 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 288 } PREFIX normal 11 delta=7626 exchange=6936 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX normal 12 delta=288 exchange=288 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX normal 13 delta=9510 exchange=0 ReductionSum { partials: 23 } PREFIX normal 14 delta=1676 exchange=888 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } RESULT injected estimated=372694 memory=MemoryPeaks { standard: 113784, interleaved: 94208, total: 169984, exchange_rows: 13648, maximum_standard_allocation: 98304 } RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } Some((ProgramCycles { total: 48984, exchange: 21336 }, MemoryUsage { standard: 0, interleaved: 0 }, 868)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 39096, exchange: 6744 }, MemoryUsage { standard: 0, interleaved: 0 }, 3280)) RESULT_OP injected Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 64 } Some((ProgramCycles { total: 49446, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7512, exchange: 6648 }, MemoryUsage { standard: 0, interleaved: 0 }, 472)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 7736, exchange: 6872 }, MemoryUsage { standard: 0, interleaved: 0 }, 688)) RESULT_OP injected ReductionSum { partials: 6 } Some((ProgramCycles { total: 6354, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 31006, exchange: 1752 }, MemoryUsage { standard: 4608, interleaved: 0 }, 2380)) RESULT_OP injected Gelu Some((ProgramCycles { total: 11370, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 49992, exchange: 7080 }, MemoryUsage { standard: 25920, interleaved: 0 }, 3208)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } Some((ProgramCycles { total: 36504, exchange: 15960 }, MemoryUsage { standard: 0, interleaved: 0 }, 652)) RESULT_OP injected Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 } Some((ProgramCycles { total: 58374, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 8520, exchange: 8088 }, MemoryUsage { standard: 0, interleaved: 0 }, 940)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 8520, exchange: 8088 }, MemoryUsage { standard: 0, interleaved: 0 }, 940)) RESULT_OP injected ReductionSum { partials: 27 } Some((ProgramCycles { total: 6954, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP injected Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 2326, exchange: 888 }, MemoryUsage { standard: 0, interleaved: 0 }, 220)) LOW_COST injected Ok(ProgramCycles { total: 260542, exchange: 67520 }) PREFIX injected 0 delta=21336 exchange=21336 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } PREFIX injected 1 delta=36006 exchange=6144 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX injected 2 delta=49446 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(2), right_inner: FromEnd(1), output_column: FromEnd(2), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 192, output_columns: 64 } PREFIX injected 3 delta=8784 exchange=6648 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX injected 4 delta=224 exchange=224 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX injected 5 delta=6354 exchange=0 ReductionSum { partials: 6 } PREFIX injected 6 delta=12867 exchange=1752 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX injected 7 delta=12219 exchange=0 Gelu PREFIX injected 8 delta=18966 exchange=7080 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX injected 9 delta=18678 exchange=15360 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } PREFIX injected 10 delta=58374 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 } PREFIX injected 11 delta=8742 exchange=8088 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX injected 12 delta=0 exchange=0 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX injected 13 delta=6954 exchange=0 ReductionSum { partials: 27 } PREFIX injected 14 delta=1592 exchange=888 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } RESULT old-only estimated=376694 memory=MemoryPeaks { standard: 77288, interleaved: 148480, total: 195584, exchange_rows: 11416, maximum_standard_allocation: 32768 } RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 41976, exchange: 4992 }, MemoryUsage { standard: 17568, interleaved: 0 }, 3100)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } Some((ProgramCycles { total: 49848, exchange: 21336 }, MemoryUsage { standard: 0, interleaved: 0 }, 976)) RESULT_OP old-only Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 } Some((ProgramCycles { total: 65742, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 9240, exchange: 8376 }, MemoryUsage { standard: 0, interleaved: 0 }, 328)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 10248, exchange: 9384 }, MemoryUsage { standard: 0, interleaved: 0 }, 472)) RESULT_OP old-only ReductionSum { partials: 8 } Some((ProgramCycles { total: 11670, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 5410, exchange: 1752 }, MemoryUsage { standard: 0, interleaved: 0 }, 580)) RESULT_OP old-only Gelu Some((ProgramCycles { total: 11370, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 49992, exchange: 7080 }, MemoryUsage { standard: 25920, interleaved: 0 }, 3208)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } Some((ProgramCycles { total: 36504, exchange: 15960 }, MemoryUsage { standard: 0, interleaved: 0 }, 652)) RESULT_OP old-only Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 } Some((ProgramCycles { total: 58374, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 8520, exchange: 8088 }, MemoryUsage { standard: 0, interleaved: 0 }, 940)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } Some((ProgramCycles { total: 8520, exchange: 8088 }, MemoryUsage { standard: 0, interleaved: 0 }, 940)) RESULT_OP old-only ReductionSum { partials: 27 } Some((ProgramCycles { total: 6954, exchange: 0 }, MemoryUsage { standard: 0, interleaved: 0 }, 0)) RESULT_OP old-only Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } Some((ProgramCycles { total: 2326, exchange: 888 }, MemoryUsage { standard: 0, interleaved: 0 }, 220)) LOW_COST old-only Ok(ProgramCycles { total: 251234, exchange: 68712 }) PREFIX old-only 0 delta=13150 exchange=4992 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX old-only 1 delta=21530 exchange=20736 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } PREFIX old-only 2 delta=65742 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 144, output_columns: 288 } PREFIX old-only 3 delta=9930 exchange=8376 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX old-only 4 delta=1440 exchange=1440 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX old-only 5 delta=11670 exchange=0 ReductionSum { partials: 8 } PREFIX old-only 6 delta=2540 exchange=1752 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX old-only 7 delta=12202 exchange=0 Gelu PREFIX old-only 8 delta=18966 exchange=7080 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } PREFIX old-only 9 delta=18402 exchange=15360 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Direct } PREFIX old-only 10 delta=58374 exchange=0 Gemm { axes: GemmAxes { left_inner: FromEnd(1), right_inner: FromEnd(2), output_column: FromEnd(1), valid_inner: None, valid_columns: None }, multiply: F16, accumulate: F32, mode: Initialize, weights: Interleaved, inner_block: 160, output_columns: 192 } PREFIX old-only 11 delta=8742 exchange=8088 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: DirectRetile, packing: Direct } PREFIX old-only 12 delta=0 exchange=0 Copy { mapping: CoordinateMapping { offsets: [1, 0, 0, 0], view: None }, policy: DirectRetile, packing: Direct } PREFIX old-only 13 delta=6954 exchange=0 ReductionSum { partials: 27 } PREFIX old-only 14 delta=1592 exchange=888 Copy { mapping: CoordinateMapping { offsets: [], view: None }, policy: StageLogicalThenTransform, packing: Staged } test planner::gemm::tests::trace_fp16_regression ... ok test result: ok. 1 passed; 0 failed; 0 ignored; 0 measured; 306 filtered out; finished in 127.41s